Well, this is not a very satisfying solution but my initial problem seems to be the sum of two problems :
- AMD's SDK 2.8 is creating problems. So for now going down to SDK 2.7 is the way to go.
- I seem to have a problem with clmagma's clAmdblas wrapper. If I call the clAmdBlas functions directly without using magma_blas, then everything works fine. This is most likely a problem on my end, but bipassing magma_blas seems to solve my problem for now.