How to use long int type in pygenn？

[OOOOOmoi]

In the generated code, it is necessary to divide the block parallel update by defining the start and ID of each group of synapses, for example:
__device__ unsigned int d_mergedPresynapticUpdateGroupStartID596[] = {4964412288, 4964975232, 4965091936, 4965207744, 4965628544, 4966049344, 4966173280, 4966297216, 4966443136, };
But when I have too many synapses it exceeds the upper limit of unsigned int type, how can I change the code to use unsigned long type?

[neworderofjamie]

What model are you trying to run here? Intended limits are:

1. Number of threads is limited to 32-bit by CUDA
2. GeNN limits number of neurons to 32-bit
3. Number of synapses can be 64-bit (although, if you are doing synapse dynamics, 1 will limit this)

However, we don't have many models that test these so 3 might be broken somewhere

[OOOOOmoi]

The model have 460w neurons and 22B synapses, I think I know what the problem is, the total number of threads exceeds 32-bit because the syn_pop. num_threads_per_spike method is used to speed up the spike processing so that the number of threads reaches 5B.
Incidentally, does pygenn have an interface for selecting device numbers? Like selectDevice?

[neworderofjamie]

Oh wow, that's a big model - I suspect increasing num_threads_per_spike won't actually help very much at that scale.

To manually select a GPU you can do this:

from pygenn.cuda_backend import DeviceSelect
...
model = GeNNModel("float", "my_model", device_select_method=DeviceSelect.MANUAL, manual_device_id: my_device_id)

[tnowotny]

Typically, the issue of a missing CUDA backend arises if the ```CUDA_PATH``` environment variable is not set or not set to the correct CUDA path.

[OOOOOmoi]

I get it, thanks.