We have noticed in the past(mid 2011) that when ADDA uses large portion of the
GPU memory, the calculation results may be incorrect. This may be caused by
reusing the GPU memory allocated to ADDA by other parts of the system (i.e. 3D
effects of the desktop). It was pronounced only on Nvidia cards, since AMD
cards enforce limit of 1/4 of GPU memory for a single object. With current ADDA
version the latter implies that ADDA may occupy less than a half of GPU memory,
decreasing the probability of memory conflicts.
This issue was not investigated systematically, but it may be caused by the
fact that some implementations (drivers) violate OpenCL standard. Thus it may
disappear by itself with future version of this drivers. Alternatively, we may
enforce the limit on single-object allocations, effectively decreasing the
available GPU memory. But it should be done only if nothing else helps.
The good first step would be to test this behavior on different hardware.
Unfortunately, currently the amount of GPU memory occupied by ADDA can be
changed minimum by a factor of 2 (due to Apple FFT routines, see issue 144),
limiting the flexibility of testing.
Original issue reported on code.google.com by yurkin on 16 Apr 2012 at 1:34
Original issue reported on code.google.com by
yurkin
on 16 Apr 2012 at 1:34