Zygote performance

[torfjelde]

Usage of Zygote.jl in Turing.jl generally comes with a really bad overhead.

I and @willtebbutt had a bit of a look at this a week ago, and it seems quite non-trivial to pin-point exactly what we need to do to make things better.

Hence, I figured we should open a discussion so we can dig into this and keep track of where we are at.

A few things that stick out immediately are:

1. @model generates conditionals that, in standard execution, usually compile way. This might not be the case for Zygote.jl and so this is likely causing type-instabilities in the model body.
2. We're using Ref for the logp field, which Zygote just treats as Ref{Any}, which again will cause type-instabilities.

I think the following is a good starting-point to determine some first points of attack:

julia> using Turing, TuringBenchmarking

julia> # Standard.
       @model demo() = x ~ Normal();

julia> benchmark_model(demo(); adbackends=[:forwarddiff, :zygote])
2-element BenchmarkTools.BenchmarkGroup:
  tags: []
  "evaluation" => 2-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "linked" => Trial(220.000 ns)
          "standard" => Trial(218.000 ns)
  "gradient" => 2-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "Turing.Essential.ForwardDiffAD{0, true}()" => 2-element BenchmarkTools.BenchmarkGroup:
                  tags: ["ForwardDiff"]
                  "linked" => Trial(315.000 ns)
                  "standard" => Trial(319.000 ns)
          "Turing.Essential.ZygoteAD()" => 2-element BenchmarkTools.BenchmarkGroup:
                  tags: ["Zygote"]
                  "linked" => Trial(133.415 μs)
                  "standard" => Trial(134.638 μs)

julia> # Circumventing the `isassumption` checks in `@model`.
       @model demo_using_addlogprob() = Turing.@addlogprob! logpdf(Normal(), __varinfo__[@varname(x)]);

julia> benchmark_model(demo_using_addlogprob(); adbackends=[:forwarddiff, :zygote], varinfo=DynamicPPL.VarInfo(demo()))
2-element BenchmarkTools.BenchmarkGroup:
  tags: []
  "evaluation" => 2-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "linked" => Trial(249.000 ns)
          "standard" => Trial(251.000 ns)
  "gradient" => 2-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "Turing.Essential.ForwardDiffAD{0, true}()" => 2-element BenchmarkTools.BenchmarkGroup:
                  tags: ["ForwardDiff"]
                  "linked" => Trial(352.000 ns)
                  "standard" => Trial(347.000 ns)
          "Turing.Essential.ZygoteAD()" => 2-element BenchmarkTools.BenchmarkGroup:
                  tags: ["Zygote"]
                  "linked" => Trial(75.975 μs)
                  "standard" => Trial(75.695 μs)

vs. completely without Turing.jl

julia> using BenchmarkTools, ForwardDiff, Zygote

julia> @benchmark $(ForwardDiff.derivative)($(Base.Fix1(logpdf, Normal())), $(randn()))
BenchmarkTools.Trial: 10000 samples with 999 evaluations.
 Range (min … max):  7.874 ns … 60.785 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     9.074 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   9.209 ns ±  1.917 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▃▁  ▂▂▂▅▅█▅▃                                               ▁
  ██▁▁█████████▇▅▃▁▁▁▃▅▅▄▃▃▄▄▄▁▄▄▃▄▁▄▁▁▁▃▁▄▄▁▁▁▁▃▁▁▃▁▁▁▁▁▄▃▄ █
  7.87 ns      Histogram: log(frequency) by time     15.5 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark $(Zygote.gradient)($(Base.Fix1(logpdf, Normal())), $(randn()))
BenchmarkTools.Trial: 10000 samples with 6 evaluations.
 Range (min … max):  5.466 μs …  1.078 ms  ┊ GC (min … max): 0.00% … 98.81%
 Time  (median):     6.285 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   6.514 μs ± 14.955 μs  ┊ GC (mean ± σ):  3.22% ±  1.40%

      ▄▂   █▆                                                 
  ▃▄▄▇██▄▄▅██▅▄▃▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂ ▃
  5.47 μs        Histogram: frequency by time        10.6 μs <

 Memory estimate: 3.38 KiB, allocs estimate: 79.

Manifest

(jl_3WWgcx) pkg> st --manifest
Status `/tmp/jl_3WWgcx/Manifest.toml`
⌃ [47edcb42] ADTypes v0.1.6
  [621f4979] AbstractFFTs v1.5.0
  [80f14c24] AbstractMCMC v4.4.2
  [7a57a42e] AbstractPPL v0.6.2
  [1520ce14] AbstractTrees v0.4.4
  [79e6a3ab] Adapt v3.6.2
  [0bf59076] AdvancedHMC v0.5.5
  [5b7e9947] AdvancedMH v0.7.5
  [576499cb] AdvancedPS v0.4.3
  [b5ca4192] AdvancedVI v0.2.4
  [dce04be8] ArgCheck v2.3.0
  [4fba245c] ArrayInterface v7.4.11
  [a9b6321e] Atomix v0.1.0
  [13072b0f] AxisAlgorithms v1.0.1
  [39de3d68] AxisArrays v0.4.7
  [198e06fe] BangBang v0.3.39
  [9718e550] Baselet v0.1.1
  [6e4b80f9] BenchmarkTools v1.3.2
  [76274a88] Bijectors v0.13.7
  [fa961155] CEnum v0.4.2
  [49dc2e85] Calculus v0.5.1
  [082447d4] ChainRules v1.54.0
  [d360d2e6] ChainRulesCore v1.16.0
  [9e997f8a] ChangesOfVariables v0.1.8
  [861a8166] Combinatorics v1.0.2
  [38540f10] CommonSolve v0.2.4
  [bbf7d656] CommonSubexpressions v0.3.0
  [34da2185] Compat v4.10.0
  [a33af91c] CompositionsBase v0.1.2
  [88cd18e8] ConsoleProgressMonitor v0.1.2
  [187b0558] ConstructionBase v1.5.4
  [a8cc5b0e] Crayons v4.1.1
  [9a962f9c] DataAPI v1.15.0
  [864edb3b] DataStructures v0.18.15
  [e2d170a0] DataValueInterfaces v1.0.0
  [244e2a9f] DefineSingletons v0.1.2
  [8bb1440f] DelimitedFiles v1.9.1
  [b429d917] DensityInterface v0.4.0
  [163ba53b] DiffResults v1.1.0
  [b552c78f] DiffRules v1.15.1
  [31c24e10] Distributions v0.25.102
  [ced4e74d] DistributionsAD v0.6.53
  [ffbed154] DocStringExtensions v0.9.3
  [fa6b7ba4] DualNumbers v0.6.8
  [366bfd00] DynamicPPL v0.23.18
  [cad2338a] EllipticalSliceSampling v1.1.0
  [4e289a0a] EnumX v1.0.4
  [e2ba6199] ExprTools v0.1.10
  [7a1cc6ca] FFTW v1.7.1
  [1a297f60] FillArrays v1.6.1
  [59287772] Formatting v0.4.2
  [f6369f11] ForwardDiff v0.10.36
  [069b7b12] FunctionWrappers v1.1.3
  [77dc65aa] FunctionWrappersWrappers v0.1.3
  [d9f16b24] Functors v0.4.5
  [0c68f7d7] GPUArrays v9.0.0
  [46192b85] GPUArraysCore v0.1.5
  [34004b35] HypergeometricFunctions v0.3.23
  [7869d1d1] IRTools v0.4.10
  [22cec73e] InitialValues v0.3.1
  [505f98c9] InplaceOps v0.3.0
  [a98d9a8b] Interpolations v0.14.7
  [8197267c] IntervalSets v0.7.7
  [3587e190] InverseFunctions v0.1.12
  [41ab1584] InvertedIndices v1.3.0
  [92d709cd] IrrationalConstants v0.2.2
  [c8e1da08] IterTools v1.8.0
  [82899510] IteratorInterfaceExtensions v1.0.0
  [692b3bcd] JLLWrappers v1.5.0
  [682c06a0] JSON v0.21.4
  [63c18a36] KernelAbstractions v0.9.8
  [5ab0869b] KernelDensity v0.6.7
  [929cbde3] LLVM v6.3.0
  [8ac3fa9e] LRUCache v1.5.0
  [b964fa9f] LaTeXStrings v1.3.0
  [50d2b5c4] Lazy v0.15.1
  [1d6d02ad] LeftChildRightSiblingTrees v0.2.0
  [6f1fad26] Libtask v0.8.6
  [6fdf6af0] LogDensityProblems v2.1.1
  [996a588d] LogDensityProblemsAD v1.6.1
  [2ab3a3ac] LogExpFunctions v0.3.26
  [e6f89c97] LoggingExtras v1.0.2
  [c7f686f2] MCMCChains v6.0.3
  [be115224] MCMCDiagnosticTools v0.3.5
  [e80e1ace] MLJModelInterface v1.9.2
  [1914dd2f] MacroTools v0.5.11
  [dbb5928d] MappedArrays v0.4.2
  [128add7d] MicroCollections v0.1.4
  [e1d29d7a] Missings v1.1.0
  [872c559c] NNlib v0.9.7
  [77ba4419] NaNMath v1.0.2
  [86f7a689] NamedArrays v0.10.0
  [c020b1a1] NaturalSort v1.0.0
  [6fe1bfb0] OffsetArrays v1.12.10
  [3bd65402] Optimisers v0.3.1
  [bac558e1] OrderedCollections v1.6.2
  [90014a1f] PDMats v0.11.19
  [69de0a69] Parsers v2.7.2
  [aea7be01] PrecompileTools v1.2.0
  [21216c6a] Preferences v1.4.1
  [08abe8d2] PrettyTables v2.2.7
  [33c8b6b6] ProgressLogging v0.1.4
  [92933f4c] ProgressMeter v1.9.0
  [1fd47b50] QuadGK v2.9.1
  [74087812] Random123 v1.6.1
  [e6cf234a] RandomNumbers v1.5.3
  [b3c3ace0] RangeArrays v0.3.2
  [c84ed2f1] Ratios v0.4.5
  [c1ae055f] RealDot v0.1.0
  [3cdcf5f2] RecipesBase v1.3.4
  [731186ca] RecursiveArrayTools v2.38.10
  [189a3867] Reexport v1.2.2
  [ae029012] Requires v1.3.0
  [37e2e3b7] ReverseDiff v1.15.1
  [79098fc4] Rmath v0.7.1
  [f2b01f46] Roots v2.0.20
  [7e49a35a] RuntimeGeneratedFunctions v0.5.12
⌅ [0bca4576] SciMLBase v1.98.1
  [c0aeaf25] SciMLOperators v0.3.6
  [30f210dd] ScientificTypesBase v3.0.0
  [efcf1570] Setfield v1.1.1
  [ce78b400] SimpleUnPack v1.1.0
  [a2af1166] SortingAlgorithms v1.1.1
  [dc90abb0] SparseInverseSubset v0.1.1
  [276daf66] SpecialFunctions v2.3.1
  [171d559e] SplittablesBase v0.1.15
  [90137ffa] StaticArrays v1.6.5
  [1e83bf80] StaticArraysCore v1.4.2
  [64bff920] StatisticalTraits v3.2.0
  [82ae8749] StatsAPI v1.7.0
  [2913bbd2] StatsBase v0.34.2
  [4c63d2b9] StatsFuns v1.3.0
  [892a3eda] StringManipulation v0.3.4
  [09ab397b] StructArrays v0.6.16
  [2efcf032] SymbolicIndexingInterface v0.2.2
  [3783bdb8] TableTraits v1.0.1
  [bd369af6] Tables v1.11.0
  [5d786b92] TerminalLoggers v0.1.7
  [9f7883ad] Tracker v0.2.27
  [28d57a85] Transducers v0.4.78
  [410a4b4d] Tricks v0.1.7
  [781d530d] TruncatedStacktraces v1.4.0
  [fce5fe82] Turing v0.29.1
  [0db1332d] TuringBenchmarking v0.3.2
  [013be700] UnsafeAtomics v0.2.1
  [d80eeb9a] UnsafeAtomicsLLVM v0.1.3
  [efce3f68] WoodburyMatrices v0.5.5
  [e88e6eb3] Zygote v0.6.65
  [700de1a5] ZygoteRules v0.2.3
  [f5851436] FFTW_jll v3.3.10+0
  [1d5cc7b8] IntelOpenMP_jll v2023.2.0+0
  [dad2f222] LLVMExtra_jll v0.0.26+0
  [856f044c] MKL_jll v2023.2.0+0
  [efe28fd5] OpenSpecFun_jll v0.5.5+0
  [f50d1b31] Rmath_jll v0.4.0+0
  [0dad84c5] ArgTools v1.1.1
  [56f22d72] Artifacts
  [2a0f44e3] Base64
  [ade2ca70] Dates
  [8ba89e20] Distributed
  [f43a241f] Downloads v1.6.0
  [7b1f6079] FileWatching
  [9fa8497b] Future
  [b77e0a4c] InteractiveUtils
  [4af54fe1] LazyArtifacts
  [b27032c2] LibCURL v0.6.3
  [76f85450] LibGit2
  [8f399da3] Libdl
  [37e2e46d] LinearAlgebra
  [56ddb016] Logging
  [d6f4376e] Markdown
  [a63ad114] Mmap
  [ca575930] NetworkOptions v1.2.0
  [44cfe95a] Pkg v1.9.2
  [de0858da] Printf
  [9abbd945] Profile
  [3fa0cd96] REPL
  [9a3f8284] Random
  [ea8e919c] SHA v0.7.0
  [9e88b42a] Serialization
  [1a1011a3] SharedArrays
  [6462fe0b] Sockets
  [2f01184e] SparseArrays
  [10745b16] Statistics v1.9.0
  [4607b0f0] SuiteSparse
  [fa267f1f] TOML v1.0.3
  [a4e569a6] Tar v1.10.0
  [8dfed614] Test
  [cf7118a7] UUIDs
  [4ec0a83e] Unicode
  [e66e0078] CompilerSupportLibraries_jll v1.0.5+0
  [deac9b47] LibCURL_jll v7.84.0+0
  [29816b5a] LibSSH2_jll v1.10.2+0
  [c8ffd9c3] MbedTLS_jll v2.28.2+0
  [14a3606d] MozillaCACerts_jll v2022.10.11
  [4536629a] OpenBLAS_jll v0.3.21+4
  [05823500] OpenLibm_jll v0.8.1+0
  [bea87d4a] SuiteSparse_jll v5.10.1+6
  [83775a58] Zlib_jll v1.2.13+0
  [8e850b90] libblastrampoline_jll v5.8.0+0
  [8e850ede] nghttp2_jll v1.48.0+0
  [3f19e933] p7zip_jll v17.4.0+0
Info Packages marked with ⌃ and ⌅ have new versions available, but those with ⌅ are restricted by compatibility constraints from upgrading. To see why use `status --outdated -m`

[willtebbutt]

Thanks for opening this @torfjelde .

This Normal distribution result with Zygote is surprisingly bad. Investigating now.

[torfjelde]

Another question: are those forward-rules used by Zygote.jl?

[willtebbutt]

Sorry, which forward-rules?

[torfjelde]

Ah sorry, these are scalar rules; my bad!

[torfjelde]

But will this actually be faster than vectorized ForwardDiff?

[willtebbutt]

Unclear. I doubt there will be a lot in it -- they should both be 10s of ns I believe, and allocation-free.