Brian2 is an improved and partly rewritten version of Brian, the spiking neural network simulator (see http://briansimulator.org). It is currently in an alpha state, ready for testing but not for "production use".
Other
0
stars
0
forks
source link
spatialdateupdate kernel_*_integration has too many parameters #1
Using the example compartmental/rall in cuda_standalone mode on my GT 610 GPU the kernel execution fails due to CUDA error (during integration): too many resources requested for launch (when calling the *_integration kernel).
The problems seems to be the kernel register usage and the resolution is to remove (at least two) unused pointers from the kernel argument list. However, as the latter is generated in cuda_standalone/device.py we need to have a way of explicitly excluding variables from that list %DEVICE_PARAMETERS% (and for %KERNEL_VARIABLES% similarily).
@Kwartke could you have a look into this? Note that the *_combine kernel was not the error source (what we initially though).
Using the example compartmental/rall in cuda_standalone mode on my GT 610 GPU the kernel execution fails due to
CUDA error (during integration): too many resources requested for launch
(when calling the *_integration kernel). The problems seems to be the kernel register usage and the resolution is to remove (at least two) unused pointers from the kernel argument list. However, as the latter is generated incuda_standalone/device.py
we need to have a way of explicitly excluding variables from that list%DEVICE_PARAMETERS%
(and for%KERNEL_VARIABLES%
similarily).@Kwartke could you have a look into this? Note that the *_combine kernel was not the error source (what we initially though).