writing an armadillo matrix produces a transposed matrix in h5

[scontini76]

writing an armadillo matrix produces a transposed matrix in h5

[steven-varga]

Great find! Hmmm there is more to this, get to it in a sec... The top one shouldn't be happening: write - read == 0 whereas the h5dump case is not quite straightforward.

The first case is a clear bug. Thanks for letting me know! The second one is complicated: Linear algebra libraries mark the layout with transpose flag but HDF5 doesn't have this feature, leading to:

1. give up zero copy (performance) and copy each dataset from col major to row major. That is lot's of copying! Considering that the majority of scientific platforms are colmajor. In cases where not enough memory is left then it will fail
2. as a feature request: convince the HDFGroup to add a flag marking transpose

Will take a closer look when I can, until I am adding @gheber to the conversation.
best:
steve

[scontini76]

Ok, let me know if you need some tests from me.
Then I wish HDFGroup consider to add a transpose flag!
In the meanwhile I'll transpose the data before writing them, or I'll do the trick of reverting be0a191