flamegpu / flamegpu2-concurrency-benchmark Goto Github PK

A bug/limitation of flamegpu2.0.0a1 limits the number of species in some cases due to an arbitrary limit.

For now, the LargePopsFalloff benchmark has been adjusted to avoid this runtime error. Once this is fixed within FLAME GPU 2 this can be increased back to match the brute force example.

See FLAMEGPU

• std::mt19937 should be switched to std::mt19937_64 to support wider seed values.

• The seed is currently provided by CLI, which if not specified will default to the current seed. Instead, to match other benchmark models we should seed with the reptition number for reproducible benchmarks (although ideally it should be cli-overrideable, but that is not currently possible for CUDASimulations)

As the DeviceMaxed benchmark uses a fixed population, irrispective of occupancy and the device used it is poorly named.

It would be better to either rename it, or make it actually max out the device (i.e. launch the maximum number of resiident threads, which will be achieved in some but not all parts of the benchmark due to occupancy constraints such as register use in some components)

CSV files should be written into the workding directory, not ../Results which may not exist depending on where the executable is called from.

Instructions may need tweaking / the plotting script moving etc, and more assumptions may need changing.

Ideally this should also not overwrite files taht exist already without a flag / user input, and not just append to files.

The large poopulation falloff benchmaks each require ~ 42 minutes of RTC compilation time in total (2 kernels * ~5s per kernel * 257 species).

This accounts for a suprisingly large amount of the total benchmark runtime.

Switching to non-RTC would remove this cost, but increase actual runtime. For spatial this change will be relatively minimal. Bruteforce would see a more significant increase (without other planned optimistaions such as shared mem curve)

Several files are commit into the build directory, this should not be the case.

Falloff experiment is not described in the readme. Either should be included or removed and data purged.