Locking Conflicts

[SimonDanisch]

Using 3 Atomics per pixel creates a ton of locking conflicts:

# Master
5.307263 seconds (251.33 M allocations: 15.794 GiB, 53.87% gc time, 356360 lock conflicts)

Not sure Immediately how to improve the situation, I guess one will need some tiling + reducers?

[anicusan]

Would using ImplicitBVH.jl be useful here? The accel directory seems to implement something fairly close to it.

[lazarusA]

@SimonDanisch I see that your work here is pretty similar the link shared. Maybe @anicusan, you could take a look and see what could be ported/removed/added, etc... The makie integration is nice, however, last time I check, those allocations at the top were really just a few and the gpu integration, only Metal was failing 😢 .

[SimonDanisch]

I should have addes some context ;) The main root of lock conflict came from the SPPMIntegrator which creates one lock per pixel.
The current BVH with multi threading via KernelAbstractions has few or zero lock conflicts.