Revisit blocksize optimisation

[genn-team]

Revisit blocksize optimisation

[tnowotny]

When running a systematic test of performance for different block sizes, the chosen model size (x in the figures), is almost always near-optimal. The model size is the number of neurons and the model is the COBAHH example. The axes are neuron kernel block sizes and synapse kernel block sizes, 32, 64, 96, 128, 256, 512, 1024. Colors are log of runtime.

model size 200

model size 400

model size 1000

model size 2000

model size 4000

model size 8000

model size 16000

model size 32000

model size 64000

model size 128000

model size 256000

model size 512000