FluxML/Zygote.jl

fastmath support

Open

#90 创建于 2019年3月6日

在 GitHub 查看
 (2 评论) (0 反应) (0 负责人)Julia (1,568 star) (220 fork)batch import
help wanted

描述

julia> foo(x) = 2.0x
foo (generic function with 1 method)

julia> bar(x) = @fastmath 2.0x
bar (generic function with 1 method)

julia> Zygote.derivative(foo, 1e6)
2.0

julia> Zygote.derivative(bar, 1e6)
ERROR: Non-differentiable function IntrinsicFunction
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] (::Zygote.Pullback{Tuple{Core.IntrinsicFunction,Float64,Float64},Tuple{Core.IntrinsicFunction}})(::Int8) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface2.jl:0
 [3] mul_fast at ./fastmath.jl:163 [inlined]
 [4] (::Zygote.Pullback{Tuple{typeof(Base.FastMath.mul_fast),Float64,Float64},Tuple{Zygote.Pullback{Tuple{Core.IntrinsicFunction,Float64,Float64},Tuple{Core.IntrinsicFunction}}}})(::Int8) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface2.jl:0
 [5] bar at ./REPL[55]:1 [inlined]
 [6] (::Zygote.Pullback{Tuple{typeof(bar),Float64},Tuple{getfield(Zygote, Symbol("##265#back#163")){getfield(Zygote, Symbol("#back#162")){:mul_fast,Zygote.Context,Module,typeof(Base.FastMath.mul_fast)}},Zygote.Pullback{Tuple{typeof(Base.FastMath.mul_fast),Float64,Float64},Tuple{Zygote.Pullback{Tuple{Core.IntrinsicFunction,Float64,Float64},Tuple{Core.IntrinsicFunction}}}}}})(::Int8) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface2.jl:0
 [7] (::getfield(Zygote, Symbol("##73#74")){Zygote.Pullback{Tuple{typeof(bar),Float64},Tuple{getfield(Zygote, Symbol("##265#back#163")){getfield(Zygote, Symbol("#back#162")){:mul_fast,Zygote.Context,Module,typeof(Base.FastMath.mul_fast)}},Zygote.Pullback{Tuple{typeof(Base.FastMath.mul_fast),Float64,Float64},Tuple{Zygote.Pullback{Tuple{Core.IntrinsicFunction,Float64,Float64},Tuple{Core.IntrinsicFunction}}}}}}})(::Int8) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface.jl:38
 [8] gradient(::Function, ::Float64) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface.jl:44
 [9] derivative(::typeof(bar), ::Float64) at /home/chriselrod/.julia/packages/Zygote/wYgcz/src/compiler/interface.jl:47
 [10] top-level scope at REPL[57]:1

@fastmath is convenient for enabling the autovectorizer and contractions (ie, fma instructions) on unrolled expressions. Would be nice to add mul_fast and friends.

贡献者指南

技术栈
无
领域
machine learning
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
4
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
3-5 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Experience with JuliaUnderstanding of Zygote.jl ADKnowledge of fast math intrinsics
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
20
研究方向
Investigate how Zygote currently handles `mul fast` and other fast math operations. The error occurs because `Core.IntrinsicFunction` is not differentiable. Look at the file `src/compiler/interface2.jl` for the pullback definition. Consider adding rules for `mul fast`, `add fast`, etc., similar to how other arithmetic operations are handled. Check for any existing discussions or PRs in the repository.