AgentSkillsCN

julia-performance-tips

在编写或优化 Julia 代码时，可运用 Julia 性能优化技巧。当您需要优化函数、排查性能问题、从性能角度审视代码，或用户就 Julia 性能、类型稳定性、内存分配或代码优化等问题向您咨询时，均可参考本指南。

SKILL.md
--- frontmatter
name: julia-performance-tips
description: Apply Julia performance optimization techniques when writing or optimizing Julia code. Use when optimizing functions, diagnosing performance issues, reviewing code for performance, or when the user asks about Julia performance, type stability, memory allocation, or code optimization.
disable-model-invocation: true

Julia Performance Tips

Essential performance optimization guidelines for Julia code. Reference: https://docs.julialang.org/en/v1/manual/performance-tips/

Core Principles

Functions and Globals

  • •Put performance-critical code in functions - code inside functions runs faster than top-level code
  • •Avoid untyped global variables - use const for globals, or pass as function arguments
  • •Break functions into multiple definitions - prefer f(x::Vector) = ... over if isa(x, Vector) ... end

Type Stability

  • •Write type-stable functions - return consistent types: use zero(x) not 0, oneunit(x) not 1
  • •Avoid changing variable types - initialize with correct type: x::Float64 = 1 not x = 1 then x /= ...
  • •Use function barriers - separate type-unstable setup from type-stable computation

Type Annotations

  • •Avoid abstract type parameters - prefer Vector{Float64} over Vector{Real}
  • •Use parametric types for struct fields - struct MyType{T} a::T end not struct MyType a::AbstractFloat end
  • •Annotate values from untyped locations - x = a[1]::Int32 when working with Vector{Any}
  • •Force specialization when needed - f(t::Type{T}) where T not f(t::Type) for Type, Function, Vararg

Memory Management

  • •Pre-allocate outputs - use in-place functions f!(out, args...) and pre-allocate out
  • •Use views for slices - @views or view() instead of array[1:5, :] when possible
  • •Fuse vectorized operations - @. 3x^2 + 4x fuses into single loop, 3x.^2 + 4x creates temporaries
  • •Unfuse when recomputing - if broadcast recomputes constant values, pre-compute: let s = sqrt.(d); x ./= s end
  • •Access arrays column-major - inner loop should vary first index: for col, row not for row, col
  • •Copy irregular views when beneficial - copying non-contiguous views can speed up repeated operations

Closures

  • •Type-annotate captured variables - r::Int = r0 in closure scope
  • •Use let blocks - f = let r = r; x -> x * r end avoids boxing
  • •Use @__FUNCTION__ for recursive closures - (@__FUNCTION__)(n-1) instead of fib(n-1)

Advanced Types

  • •Use Val for compile-time values - f(::Val{N}) where N when dimension known at compile time
  • •Avoid excessive type parameters - only use values-as-parameters when processing homogeneous collections

Performance Tools

  • •@time - measure time and allocations (ignore first run, it's compilation)
  • •@code_warntype - find type instabilities (red = non-concrete types)
  • •@allocated - measure memory allocations
  • •Profiling - use Profile.jl or ProfileView.jl for bottlenecks
  • •JET.jl - static analysis for performance issues
  • •--track-allocation=user - find allocation sources

Performance Annotations

  • •@inbounds - disable bounds checking (use with caution)
  • •@fastmath - allow floating-point optimizations (may change results)
  • •@simd - promise independent loop iterations (experimental, use carefully)

Miscellaneous

  • •Avoid unnecessary arrays - x+y+z not sum([x,y,z])
  • •Use abs2 for complex numbers - abs2(z) not abs(z)^2
  • •Use div, fld, cld - not trunc(x/y), floor(x/y), ceil(x/y)
  • •Fix deprecation warnings - they add lookup overhead
  • •Avoid string interpolation for I/O - println(file, a, " ", b) not println(file, "$a $b")
  • •Use LazyString for conditional strings - lazy"..." for error paths
  • •Set OPENBLAS_NUM_THREADS=1 when using JULIA_NUM_THREADS>1 for multithreaded code

Package Performance

  • •Use PrecompileTools.jl - reduce time-to-first-execution
  • •Minimize dependencies - use package extensions for optional features
  • •Avoid heavy __init__() - minimize compilation in initialization
  • •Use @time_imports - diagnose slow package loading