[BUG] ORC reader produces different output than Pandas ORC reader reading a specific ORC file.

[firestarman]

This can be reproduced by reading the attached ORC file ( struct_orc.log ).

Starting at line 995, you can find the data in nested column s1.sa are different.

>>> pandas.read_orc("/data/struct_845472261.orc")
               a             b                     s1                             s2
0    326215455.0  9.646704e+08   {'sa': 2056831253.0}  {'sa': {'ssa': 1140062029.0}}
1    326215455.0 -1.123178e+09   {'sa': 1735783038.0}                           None
2    326215455.0           NaN   {'sa': -335532064.0}  {'sa': {'ssa': 1998862860.0}}
3    326215455.0 -6.566856e+08  {'sa': -1463792165.0}                   {'sa': None}
4    326215455.0 -1.943089e+09   {'sa': 1762505979.0}  {'sa': {'ssa': -521321651.0}}
..           ...           ...                    ...                            ...
995  289956431.0  1.954978e+09    {'sa': 281568135.0}  {'sa': {'ssa': -873056332.0}}
996  289956431.0 -1.028878e+09   {'sa': -899673285.0}  {'sa': {'ssa': 1353167471.0}}
997  289956431.0  1.827136e+09   {'sa': -950538642.0}                           None
998  289956431.0  2.147484e+09  {'sa': -1024796596.0}  {'sa': {'ssa': 1126348031.0}}
999  289956431.0 -1.223061e+08                   None  {'sa': {'ssa': 1008923253.0}}

[1000 rows x 4 columns]
>>> 
>>> import cudf
>>> cudf.read_orc("/data/struct_845472261.orc")
             a            b                    s1                          s2
0    326215455    964670395   {'0': 2056831253.0}  {'0': {'0': 1140062029.0}}
1    326215455  -1123177776   {'0': 1735783038.0}                        None
2    326215455         <NA>   {'0': -335532064.0}  {'0': {'0': 1998862860.0}}
3    326215455   -656685553  {'0': -1463792165.0}                 {'0': None}
4    326215455  -1943088595   {'0': 1762505979.0}  {'0': {'0': -521321651.0}}
..         ...          ...                   ...                         ...
995  289956431   1954977719  {'0': -1219859255.0}  {'0': {'0': 2013006809.0}}
996  289956431  -1028878225   {'0': 1139661239.0}   {'0': {'0': 908256279.0}}
997  289956431   1827136073    {'0': 871691692.0}                        None
998  289956431   2147483647    {'0': 720591207.0}  {'0': {'0': 1773585916.0}}
999  289956431   -122306082                  None  {'0': {'0': 1801980339.0}}

[1000 rows x 4 columns]
>>> 

Similiar to the file mentioned in the issue #8878, this file "struct_orc.log" is also generated by Spark rapids. And both Pandas and Spark can read it correctly.

[firestarman]

What's more, the nested column names seem to be dropped according to the output, instead "0" is used as the name for all the nested columns. I am not sure this is correct.

[vuule]

Some observations that might be useful:
in the s1 column, cudf turns the element in row 90 into null, and moves null element from row 99 to row 97. The data stream seems to be parsed correctly, just the null stream seems incorrect.

@rgsl888prabhu , there was an issue with the null masks in the writer when row groups stopped being divisible by 8. Some assumptions about byte alignment broke down and I had to adjust the indexing. Could something similar be happening here?
Edit: Just realized that this case is accounted for in the code.

[rgsl888prabhu]

Yeah, we are using atomic OR, but not sure where we are messing up.