https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121760
--- Comment #20 from Benjamin Schulz <schulz.benjamin at googlemail dot com> --- hi Jacub, thats exactly what worries me, and for this reason, my code has a flag, which says it should often not map the dpdata field at all, and then will map only strides and extents, while the dpdata field was allocated as a devicepointer with target_alloc... But for a matrix, strides and extents only have rank 2, or size 2 and only these should get mapped if at all repeatedly, and of course then entirely... But i will look again at my code to see whether Ive forgotten to set that flag somewhere that it should not map the dpdata field... And also, aehm, the error should then be there if i put the same code of the naive matrix multiplication in to the strassen multiply function. But then the error vanishes... But I will look again at my code.. and see if i can offload the precompiled code f the problem persists...
