There was an error in conjugate-transpose part (her=true) of matvec product for
a long time (at least from release 1.1). This part is used only in CGNR
iterative solver. Moreover, it was negligible for a sphere, so a different
shape (with different volume fraction of a computational grid) is required to
notice it (that's why it was not noticed by automatic tests).
For example, simple test is 'adda_ocl.exe -iter cgnr -shape box', which causes
convergence failure.
Original issue reported on code.google.com by yurkin on 28 May 2014 at 11:42
Original issue reported on code.google.com by
yurkin
on 28 May 2014 at 11:42