* Remove all occurencies of threading::memory_order* in multi_array_chunked.hxx
(or equivalently, replace them with threading::memory_order_seq_cst). This will
force all atomic accesses to use sequentially consistent ordering, the most
strict possibility.
That worked out ... much thanks!
- First of all, there was a little mistake, that was the wrong log file.
Unfortunately, the
problem with test_multiarray_chunked resp. multi_array_chunked.hxx on
ppc64el/le remains also
with changing `memory_order_release/acquire` to `memory_order_seq_cst`.
bummer!
I'm further experimenting with VIGRA_SINGLE_THREADED but no breakthrough so far
...
Looking at the log again, it occured to me that the error might not be
related to multi-threading at all. Possibly, all failing tests are on
ChunkedArrayCompressed<T> with T being float and double. Can you please
check this? If this is true, the error might go away be changing the
compression algorithm in test_chunked.cxx line 121 from LZ4 into
ZLIB_FAST or NO_COMPRESSION. If this works, there might be a byte order
problem in the LZ4 algorithm. Endianess is decided in
<vigra>/src/impex/lz4.c lines 65 to 79. Please check if the correct
endianess for your platform is retrieved there, and possibly propose a
fix if it isn't.
Best
Ulli
