Add PrecompileTools

[vchuravy]

No description provided.

[pxl-th]

Temporary removed AMDGPU to make it run on Julia 1.11 (needs update to LLVM 8 for GPUCompiler master).

You can run julia benchmark/pipeline.jl from Nerf.jl directory to run the whole pipeline.

Without PrecompileTools:

Trainer benchmark
 17.861788 seconds (30.04 M allocations: 1.492 GiB, 2.76% gc time, 74.64% compilation time: <1% of which was recompilation)
 39.232965 seconds (4.11 M allocations: 124.968 MiB, 1.58% gc time)
Renderer benchmark
 13.524578 seconds (10.00 M allocations: 336.965 MiB, 3.08% gc time, 14.66% compilation time)
 51.170871 seconds (33.40 M allocations: 776.366 MiB, 3.85% gc time)

With PrecompileTools:

Trainer benchmark
  4.884491 seconds (1.24 M allocations: 112.379 MiB, 19.31% gc time, 7.50% compilation time)
 36.660421 seconds (4.10 M allocations: 124.746 MiB, 0.70% gc time)
Renderer benchmark
 11.767535 seconds (7.21 M allocations: 196.629 MiB, 1.99% gc time, 0.43% compilation time)
 52.402654 seconds (34.08 M allocations: 794.252 MiB, 2.86% gc time)

[vchuravy]

I wonder if enable the disk cache would do something, but it is odd that the trainer benchmark still spends 70% in compilation

[pxl-th]

Oh, indeed :) Now it is faster (updated above post). Benchmark consist of:

@time trainer_benchmark(trainer, 10)
@time trainer_benchmark(trainer, 1000)

Where first line is to precompile all kernels mainly. So the time went down from ~18 seconds to ~5 seconds.