Use Bumper.jl to reduce allocations in bulk sampling

[ameligrana]

No description provided.

[ameligrana]

Another dependency, but still lightweight and reasonably stable

[github-actions]

Benchmark Results

	3a191dc	eaa8e9d	3a191dc / eaa8e9d
TTFX excluding time to load	1.02 ± 0 s	0.932 ± 0 s	1.09 ± 0,1.1 ± 0,1.09 ± 0
code size in bytes	1.51e+04 ± 0 h	1.51e+04 ± 0 h	1 ± 0
code size in lines	566 ± 0 h	564 ± 0 h	1 ± 0
code size in syntax nodes	4.08e+03 ± 0 h	4.08e+03 ± 0 h	1 ± 0
constructor n=100 σ=0.1	5.88 ± 0.2 μs	5.87 ± 0.15 μs	0.984 ± 0.042,0.973 ± 0.1,1 ± 0.044
constructor n=100 σ=1.0	6.27 ± 0.16 μs	6.32 ± 0.26 μs	0.973 ± 0.054,0.988 ± 0.061,0.992 ± 0.049
constructor n=100 σ=10.0	6.84 ± 0.38 μs	6.7 ± 0.17 μs	0.999 ± 0.055,0.989 ± 0.049,1.02 ± 0.063
constructor n=100 σ=100.0	0.0428 ± 0.0031 ms	10.5 ± 7.9 μs	4.18 ± 1.9,0.994 ± 0.89,4.06 ± 3.1
constructor n=1000 σ=0.1	0.0517 ± 0.0044 ms	0.042 ± 0.0013 ms	0.979 ± 0.043,1.02 ± 0.044,1.23 ± 0.11
constructor n=1000 σ=1.0	0.0448 ± 0.001 ms	0.0441 ± 0.001 ms	0.983 ± 0.04,1.01 ± 0.059,1.02 ± 0.034
constructor n=1000 σ=10.0	0.0517 ± 0.001 ms	0.056 ± 0.0047 ms	0.984 ± 0.033,1 ± 0.038,0.922 ± 0.079
constructor n=1000 σ=100.0	0.0601 ± 0.0012 ms	0.0597 ± 0.00082 ms	0.985 ± 0.03,0.996 ± 0.029,1.01 ± 0.024
constructor n=10000 σ=0.1	0.433 ± 0.03 ms	0.418 ± 0.024 ms	0.982 ± 0.074,1.02 ± 0.071,1.04 ± 0.093
constructor n=10000 σ=1.0	0.433 ± 0.025 ms	0.431 ± 0.026 ms	0.97 ± 0.065,1.01 ± 0.08,1 ± 0.083
constructor n=10000 σ=10.0	0.458 ± 0.028 ms	0.521 ± 0.03 ms	0.952 ± 0.1,1.15 ± 0.78,0.879 ± 0.073
constructor n=10000 σ=100.0	0.583 ± 0.026 ms	0.584 ± 0.034 ms	0.989 ± 0.075,0.999 ± 0.082,0.998 ± 0.073
delete ∘ rand n=100 σ=0.1	4.48 ± 0.19 μs	4.47 ± 0.2 μs	1 ± 0.06,1.01 ± 0.061,1 ± 0.062
delete ∘ rand n=100 σ=1.0	4.75 ± 0.2 μs	4.7 ± 0.19 μs	0.996 ± 0.058,1.01 ± 0.056,1.01 ± 0.059
delete ∘ rand n=100 σ=10.0	4.89 ± 0.18 μs	4.87 ± 0.18 μs	0.992 ± 0.052,1.01 ± 0.052,1 ± 0.052
delete ∘ rand n=100 σ=100.0	7.27 ± 0.47 μs	7.28 ± 0.45 μs	0.989 ± 0.088,0.989 ± 0.092,0.999 ± 0.089
delete ∘ rand n=1000 σ=0.1	0.0445 ± 0.00075 ms	0.0443 ± 0.00093 ms	1 ± 0.027,1.01 ± 0.028,1.01 ± 0.027
delete ∘ rand n=1000 σ=1.0	0.0484 ± 0.001 ms	0.0479 ± 0.00096 ms	0.996 ± 0.029,1.01 ± 0.03,1.01 ± 0.029
delete ∘ rand n=1000 σ=10.0	0.0494 ± 0.00086 ms	0.0491 ± 0.0011 ms	0.992 ± 0.025,1 ± 0.027,1 ± 0.028
delete ∘ rand n=1000 σ=100.0	0.0541 ± 0.001 ms	0.0541 ± 0.001 ms	0.992 ± 0.024,1 ± 0.027,0.999 ± 0.026
delete ∘ rand n=10000 σ=0.1	0.481 ± 0.012 ms	0.482 ± 0.019 ms	0.971 ± 0.12,1.01 ± 0.037,0.998 ± 0.047
delete ∘ rand n=10000 σ=1.0	0.515 ± 0.012 ms	0.518 ± 0.014 ms	0.995 ± 0.03,1.01 ± 0.035,0.995 ± 0.036
delete ∘ rand n=10000 σ=10.0	0.522 ± 0.016 ms	0.522 ± 0.014 ms	0.988 ± 0.043,1.01 ± 0.038,1 ± 0.041
delete ∘ rand n=10000 σ=100.0	0.52 ± 0.013 ms	0.514 ± 0.014 ms	0.985 ± 0.043,1.01 ± 0.045,1.01 ± 0.037
empty constructor	1.89 ± 0.16 μs	2.11 ± 0.32 μs	0.928 ± 0.2,0.92 ± 0.26,0.899 ± 0.16
intermixed_h n=100 σ=0.1	11.4 ± 1.2 μs	14.8 ± 1.4 μs	0.978 ± 0.11,0.984 ± 0.12,0.77 ± 0.11
intermixed_h n=100 σ=1.0	11.7 ± 1.1 μs	11.7 ± 1.1 μs	0.986 ± 0.13,0.99 ± 0.15,1 ± 0.13
intermixed_h n=100 σ=10.0	11.5 ± 1.7 μs	11.5 ± 2 μs	0.988 ± 0.23,0.994 ± 0.33,0.998 ± 0.23
intermixed_h n=100 σ=100.0	12.3 ± 1.2 μs	12.4 ± 1.4 μs	0.974 ± 0.15,1 ± 0.15,0.99 ± 0.15
intermixed_h n=1000 σ=0.1	0.108 ± 0.0089 ms	0.115 ± 0.0091 ms	0.979 ± 0.1,0.937 ± 0.13,0.941 ± 0.11
intermixed_h n=1000 σ=1.0	0.111 ± 0.0092 ms	0.109 ± 0.0085 ms	0.978 ± 0.11,0.985 ± 0.11,1.02 ± 0.12
intermixed_h n=1000 σ=10.0	0.104 ± 0.0096 ms	0.107 ± 0.01 ms	0.974 ± 0.11,1.03 ± 0.13,0.973 ± 0.13
intermixed_h n=1000 σ=100.0	0.113 ± 0.013 ms	0.114 ± 0.014 ms	0.975 ± 0.15,0.989 ± 0.15,0.99 ± 0.17
intermixed_h n=10000 σ=0.1	1.2 ± 0.19 ms	1.13 ± 0.19 ms	1.02 ± 0.22,1.03 ± 0.22,1.06 ± 0.25
intermixed_h n=10000 σ=1.0	1.2 ± 0.2 ms	1.21 ± 0.21 ms	0.956 ± 0.2,0.968 ± 0.21,0.988 ± 0.23
intermixed_h n=10000 σ=10.0	1.09 ± 0.17 ms	1.08 ± 0.13 ms	0.974 ± 0.2,1.03 ± 0.27,1.01 ± 0.2
intermixed_h n=10000 σ=100.0	1.14 ± 0.18 ms	1.11 ± 0.18 ms	1 ± 0.21,0.978 ± 0.2,1.03 ± 0.23
pathological 1	0.0547 ± 0.00032 μs	0.0468 ± 0.0011 μs	1.01 ± 0.0094,1.01 ± 0.0085,1.17 ± 0.028
pathological 1′	0.162 ± 0.0016 μs	0.18 ± 0.0017 μs	1.03 ± 0.031,1.01 ± 0.013,0.902 ± 0.012
pathological 2	0.0749 ± 0.00016 μs	0.0636 ± 0.00031 μs	1.01 ± 0.0088,1.01 ± 0.0054,1.18 ± 0.0062
pathological 2′	0.18 ± 0.0016 μs	0.171 ± 0.0016 μs	1.09 ± 0.022,1 ± 0.013,1.05 ± 0.013
pathological 2′′	0.21 ± 0.0083 μs	0.206 ± 0.005 μs	1.08 ± 0.041,1.04 ± 0.11,1.02 ± 0.047
pathological 3	16.9 ± 0.24 ns	17 ± 0.23 ns	1 ± 0.019,1 ± 0.022,0.997 ± 0.02
pathological 4	0.0743 ± 0.00018 μs	0.063 ± 0.00029 μs	1.1 ± 0.0076,1.01 ± 0.0066,1.18 ± 0.0061
pathological 4′	0.185 ± 0.0017 μs	0.185 ± 0.0014 μs	1.06 ± 0.019,0.986 ± 0.014,0.999 ± 0.012
pathological 4′′	0.22 ± 0.0047 μs	0.217 ± 0.004 μs	1.09 ± 0.055,1.03 ± 0.051,1.02 ± 0.029
pathological 5a	0.0542 ± 0.00011 μs	0.0453 ± 0.00025 μs	1.01 ± 0.0077,1.01 ± 0.0071,1.2 ± 0.007
pathological 5b	0.0542 ± 0.00011 μs	0.0453 ± 0.00026 μs	1.01 ± 0.0079,1.01 ± 0.0077,1.2 ± 0.0074
pathological 5b′	0.325 ± 0.0032 μs	0.329 ± 0.0033 μs	1.02 ± 0.015,1.02 ± 0.015,0.989 ± 0.014
pathological 5b′′	0.336 ± 0.0084 μs	0.334 ± 0.0062 μs	0.991 ± 0.031,1 ± 0.033,1.01 ± 0.031
pathological large compaction (133380-op)	15.2 ± 0.12 ms	15.1 ± 0.24 ms	1.03 ± 0.03,1.01 ± 0.036,1 ± 0.018
pathological medium compaction (1254-op)	0.141 ± 0.096 ms	0.102 ± 0.0029 ms	0.985 ± 0.32,1.06 ± 0.24,1.38 ± 0.94
pathological old compaction (6-op)	0.211 μs	0.28 μs	1.04,1.01,0.755
pathological small compaction (18-op)	0.922 ± 0.022 μs	0.919 ± 0.013 μs	1.01 ± 0.022,1.01 ± 0.019,1 ± 0.028
pathological tiny compaction (6-op)	0.296 ± 0.0021 μs	0.285 ± 0.011 μs	1.05 ± 0.024,1.03 ± 0.014,1.04 ± 0.041
sample (bulk) n=1000 k=10000 σ=1	0.166 ± 0.02 ms	0.167 ± 0.021 ms	1 ± 0.19,0.995 ± 0.18,0.997 ± 0.17
sample (bulk) n=1000 k=10000 σ=100	0.101 ± 0.034 ms	0.103 ± 0.037 ms	1 ± 0.51,0.977 ± 0.5,0.986 ± 0.49
sample (bulk) n=1000 k=1000000 σ=1	16.5 ± 1.8 ms	16.3 ± 2.2 ms	0.987 ± 0.17,1.03 ± 0.19,1.02 ± 0.18
sample (bulk) n=1000 k=1000000 σ=100	8.12 ± 3 ms	8.2 ± 3.4 ms	1.05 ± 0.66,0.97 ± 0.53,0.991 ± 0.55
sample (bulk) n=1000000 k=10000 σ=1	0.337 ± 0.015 ms	0.368 ± 0.032 ms	0.987 ± 0.14,0.999 ± 0.14,0.916 ± 0.09
sample (bulk) n=1000000 k=10000 σ=100	0.168 ± 0.008 ms	0.155 ± 0.043 ms	1.15 ± 0.41,1.09 ± 0.39,1.08 ± 0.3
sample (bulk) n=1000000 k=1000000 σ=1	21.8 ± 0.64 ms	21.8 ± 1.4 ms	1.06 ± 0.032,1.04 ± 0.056,1 ± 0.069
sample (bulk) n=1000000 k=1000000 σ=100	6.86 ± 1.5 ms	8.67 ± 2.9 ms	0.818 ± 0.36,1.09 ± 0.65,0.791 ± 0.32
sample n=100 σ=0.1	25.4 ± 0.63 ns	25.5 ± 0.71 ns	1 ± 0.035,1 ± 0.036,0.997 ± 0.037
sample n=100 σ=1.0	29.5 ± 2.2 ns	29.6 ± 2.4 ns	0.998 ± 0.11,0.999 ± 0.11,0.999 ± 0.11
sample n=100 σ=10.0	19.3 ± 4.5 ns	19.6 ± 4.6 ns	1.01 ± 0.35,1.01 ± 0.35,0.986 ± 0.33
sample n=100 σ=100.0	16.9 ± 3.9 ns	17.1 ± 3.9 ns	1.01 ± 0.32,1.01 ± 0.33,0.989 ± 0.32
sample n=1000 σ=0.1	23.5 ± 6.3 ns	23.7 ± 6.1 ns	1.01 ± 0.35,1.01 ± 0.29,0.991 ± 0.37
sample n=1000 σ=1.0	0.0321 ± 0.0023 μs	0.0324 ± 0.0024 μs	1 ± 0.099,1 ± 0.098,0.992 ± 0.1
sample n=1000 σ=10.0	20.8 ± 5.6 ns	20.4 ± 5.3 ns	0.995 ± 0.37,1 ± 0.38,1.02 ± 0.38
sample n=1000 σ=100.0	16.9 ± 4 ns	17.5 ± 4.1 ns	1.01 ± 0.33,1.01 ± 0.34,0.968 ± 0.33
sample n=10000 σ=0.1	0.0318 ± 0.0011 μs	0.0318 ± 0.0013 μs	1 ± 0.083,0.996 ± 0.057,0.999 ± 0.054
sample n=10000 σ=1.0	0.0342 ± 0.0015 μs	0.0344 ± 0.0011 μs	1.01 ± 0.071,1.01 ± 0.05,0.994 ± 0.055
sample n=10000 σ=10.0	22.3 ± 6.6 ns	22.1 ± 6.2 ns	1.03 ± 0.42,1 ± 0.4,1.01 ± 0.41
sample n=10000 σ=100.0	17.5 ± 4.2 ns	18 ± 4.1 ns	1.01 ± 0.34,1.04 ± 0.32,0.973 ± 0.32
summarysize n=100 σ=0.1	1.19e+05 ± 0 h	1.19e+05 ± 0 h	1 ± 0
summarysize n=100 σ=1.0	1.19e+05 ± 0 h	1.19e+05 ± 0 h	1 ± 0
summarysize n=100 σ=10.0	1.19e+05 ± 0 h	1.19e+05 ± 0 h	1 ± 0
summarysize n=100 σ=100.0	1.19e+05 ± 0 h	1.19e+05 ± 0 h	1 ± 0
summarysize n=1000 σ=0.1	1.52e+05 ± 0 h	1.52e+05 ± 0 h	1 ± 0
summarysize n=1000 σ=1.0	1.52e+05 ± 0 h	1.52e+05 ± 0 h	1 ± 0
summarysize n=1000 σ=10.0	1.52e+05 ± 0 h	1.52e+05 ± 0 h	1 ± 0
summarysize n=1000 σ=100.0	1.52e+05 ± 0 h	1.52e+05 ± 0 h	1 ± 0
summarysize n=10000 σ=0.1	1.13e+06 ± 0 h	1.13e+06 ± 0 h	1 ± 0
summarysize n=10000 σ=1.0	1.13e+06 ± 0 h	1.13e+06 ± 0 h	1 ± 0
summarysize n=10000 σ=10.0	1.13e+06 ± 0 h	1.13e+06 ± 0 h	1 ± 0
summarysize n=10000 σ=100.0	1.13e+06 ± 0 h	1.13e+06 ± 0 h	1 ± 0
update ∘ rand n=100 σ=0.1	0.0857 ± 0.002 μs	0.0821 ± 0.0021 μs	1.03 ± 0.043,1.02 ± 0.036,1.04 ± 0.036
update ∘ rand n=100 σ=1.0	0.0918 ± 0.0024 μs	0.089 ± 0.0028 μs	1.03 ± 0.046,1.01 ± 0.042,1.03 ± 0.042
update ∘ rand n=100 σ=10.0	0.102 ± 0.0032 μs	0.0989 ± 0.0035 μs	1.02 ± 0.049,1.01 ± 0.046,1.04 ± 0.049
update ∘ rand n=100 σ=100.0	0.175 ± 0.015 μs	0.175 ± 0.016 μs	0.986 ± 0.12,0.99 ± 0.12,1 ± 0.12
update ∘ rand n=1000 σ=0.1	0.0861 ± 0.0022 μs	0.0827 ± 0.0022 μs	1.03 ± 0.04,1.02 ± 0.035,1.04 ± 0.039
update ∘ rand n=1000 σ=1.0	0.0917 ± 0.002 μs	0.0889 ± 0.0023 μs	1.02 ± 0.03,1.01 ± 0.031,1.03 ± 0.034
update ∘ rand n=1000 σ=10.0	0.0998 ± 0.0022 μs	0.0964 ± 0.0018 μs	1.02 ± 0.032,1.01 ± 0.031,1.04 ± 0.03
update ∘ rand n=1000 σ=100.0	0.172 ± 0.006 μs	0.171 ± 0.0074 μs	0.991 ± 0.053,0.994 ± 0.053,1 ± 0.056
update ∘ rand n=10000 σ=0.1	0.0928 ± 0.0013 μs	0.0905 ± 0.0013 μs	1.01 ± 0.02,1.01 ± 0.025,1.03 ± 0.021
update ∘ rand n=10000 σ=1.0	0.096 ± 0.0021 μs	0.0943 ± 0.0012 μs	1.01 ± 0.024,1.01 ± 0.014,1.02 ± 0.026
update ∘ rand n=10000 σ=10.0	0.0975 ± 0.0014 μs	0.0951 ± 0.001 μs	1.01 ± 0.024,0.999 ± 0.016,1.03 ± 0.019
update ∘ rand n=10000 σ=100.0	0.166 ± 0.0023 μs	0.166 ± 0.0024 μs	0.986 ± 0.023,0.981 ± 0.017,1 ± 0.02
time_to_load	0.0848 ± 0.0014 s	0.0834 ± 0.0013 s	1.02 ± 0.0091,1.03 ± 0.014,1.02 ± 0.023

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[ameligrana]

Seems to have a significant effect on performance, and allocations are now practically only the one of the sample, though Bumper internals scares me a bit

[ameligrana]

I'm wondering if we can use for the same purpose yours https://github.com/LilithHafner/PtrArrays.jl @LilithHafner

[ameligrana]

Tried it, seems slower than Bumper.jl when I wrap the free calls in a try-finally block, and faster if I do not, but having that block seems strongly encouraged by the readme

[ameligrana]

Actually it is really needed for safety the try-finally block, but I will open up a new PR with PtrArrays to try it anyway

[LilithHafner]

What benchmarks motivate this?
How does this impact precompile time?

[ameligrana]

What benchmarks motivate this?

All cases where k is sufficiently low e.g. in our benchmarks they are

sample (bulk) n=1000 k=10000 σ=1
sample (bulk) n=1000 k=10000 σ=100
sample (bulk) n=1000000 k=10000 σ=1
sample (bulk) n=1000000 k=10000 σ=100

actually we don't see a significant effect in our benchmarks consistently, but this is because the effect is more on the average than on the minimum (around 10-15% there).

How does this impact precompile time?

with a fresh environment on 1.11:

without Bumper:

julia> @time @eval using WeightVectors
Precompiling WeightVectors...
  4 dependencies successfully precompiled in 2 seconds
  2.040833 seconds (58.16 k allocations: 4.299 MiB, 1.00% compilation time)

with Bumper:

julia> @time @eval using WeightVectors
Precompiling WeightVectors...
  6 dependencies successfully precompiled in 2 seconds
  2.444766 seconds (75.47 k allocations: 5.830 MiB, 0.90% compilation time)