Fixed nasty race condition

[neworderofjamie]

In #376 I added support for tracking 'previous' spike times. However, this generated preNeuronResetKernel code like this:

if(lid < group->spkCnt[0]) {
    group->prevST[group->spk[lid]] = t - DT;
}
        
if(lid == 0) {
    group->spkCnt[0] = 0;
}

Which has a nasty race condition in that the spike count may have been reset by the time other blocks try and access it. Presumably, this wasn't hit in our tests or the Pavlovian conditioning model because they were small enough that the whole kernel ran simultaneously. Better performing (this adds approximately 1.5 microseconds per-timestep to the Pavlovian conditioning model), more complex solutions might be possible but for now I've just split the two stages into two separate kernels which has had the side benefit of tidying up several nasty bits of the previous implementation)

[codecov]

Codecov Report

Merging #414 (b8ad859) into master (f4240a5) will decrease coverage by 0.02%.
The diff coverage is 85.71%.

@@            Coverage Diff             @@
##           master     #414      +/-   ##
==========================================
- Coverage   87.13%   87.10%   -0.03%     
==========================================
  Files          78       78              
  Lines       15245    15309      +64     
==========================================
+ Hits        13283    13335      +52     
- Misses       1962     1974      +12     

Impacted Files	Coverage Δ
src/genn/backends/single_threaded_cpu/backend.cc	56.35% <47.82%> (-0.02%)	⬇️
src/genn/genn/code_generator/backendSIMT.cc	96.42% <74.02%> (-0.09%)	⬇️
include/genn/genn/code_generator/backendSIMT.h	98.48% <100.00%> (-0.05%)	⬇️
include/genn/genn/code_generator/groupMerged.h	91.15% <100.00%> (+0.08%)	⬆️
include/genn/genn/code_generator/modelSpecMerged.h	98.78% <100.00%> (+0.01%)	⬆️
src/genn/backends/cuda/backend.cc	87.85% <100.00%> (+0.17%)	⬆️
src/genn/backends/cuda/optimiser.cc	77.73% <100.00%> (-0.10%)	⬇️
src/genn/backends/opencl/backend.cc	91.11% <100.00%> (+0.19%)	⬆️
src/genn/genn/code_generator/generateRunner.cc	95.99% <100.00%> (+0.01%)	⬆️
src/genn/genn/code_generator/groupMerged.cc	87.04% <100.00%> (+0.14%)	⬆️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f4240a5...b8ad859. Read the comment docs.

[tnowotny]

Classic trap. Glad you spotted it. Trying to think of more efficient solutions but coming up blank atm.

[neworderofjamie]

Sadly not the cause of the bug that led me to fix it but hey 😄