Merge pull request #1072 from SciML/fastpow

Add Enzyme support for fastpow

ext/DiffEqBaseEnzymeExt.jl

@@ -1,7 +1,7 @@
 module DiffEqBaseEnzymeExt
 
 using DiffEqBase
-import DiffEqBase: value
+import DiffEqBase: value, fastpow
 using Enzyme
 import Enzyme: Const
 using ChainRulesCore
@@ -53,4 +53,6 @@ function Enzyme.EnzymeRules.reverse(config::Enzyme.EnzymeRules.RevConfigWidth{1}
     return ntuple(_ -> nothing, Val(length(args) + 4))
 end
 
-end
+Enzyme.Compiler.known_ops[typeof(DiffEqBase.fastpow)] = (:pow, 2, nothing)
+
+end 

src/fastpow.jl

@@ -51,60 +51,20 @@ const EXP2FT = (Float32(0x1.6a09e667f3bcdp-1),
     Float32(0x1.3dea64c123422p+0),
     Float32(0x1.4bfdad5362a27p+0),
     Float32(0x1.5ab07dd485429p+0))
-@inline function _exp2(x::Float32)
-    TBLBITS = UInt32(4)
-    TBLSIZE = UInt32(1 << TBLBITS)
 
-    redux = Float32(0x1.8p23) / TBLSIZE
-    P1 = Float32(0x1.62e430p-1)
-    P2 = Float32(0x1.ebfbe0p-3)
-    P3 = Float32(0x1.c6b348p-5)
-    P4 = Float32(0x1.3b2c9cp-7)
-
-    # Reduce x, computing z, i0, and k.
-    t::Float32 = x + redux
-    i0 = reinterpret(UInt32, t)
-    i0 += TBLSIZE ÷ UInt32(2)
-    k::UInt32 = unsafe_trunc(UInt32, (i0 >> TBLBITS) << 20)
-    i0 &= TBLSIZE - UInt32(1)
-    t -= redux
-    z = x - t
-    twopk = Float32(reinterpret(Float64, UInt64(0x3ff00000 + k) << 32))
-
-    # Compute r = exp2(y) = exp2ft[i0] * p(z).
-    tv = EXP2FT[i0 + UInt32(1)]
-    u = tv * z
-    tv = tv + u * (P1 + z * P2) + u * (z * z) * (P3 + z * P4)
-
-    # Scale by 2**(k>>20)
-    return tv * twopk
-end
-
-if VERSION < v"1.7.0"
-    """
-        fastpow(x::Real, y::Real) -> Float32
-    """
-    @inline function fastpow(x::Real, y::Real)
-        if iszero(x)
-            return 0.0f0
-        elseif isinf(x) && isinf(y)
-            return Float32(Inf)
-        else
-            return _exp2(convert(Float32, y) * fastlog2(convert(Float32, x)))
-        end
-    end
-else
-    """
-        fastpow(x::Real, y::Real) -> Float32
-    """
-    @inline function fastpow(x::Real, y::Real)
-        if iszero(x)
-            return 0.0f0
-        elseif isinf(x) && isinf(y)
-            return Float32(Inf)
-        else
-            return @fastmath exp2(convert(Float32, y) * fastlog2(convert(Float32, x)))
-        end
+"""
+    fastpow(x::T, y::T) where {T} -> float(T)
+    Trips through Float32 for performance.
+"""
+@inline function fastpow(x::T, y::T) where {T}
+    outT = float(T)
+    if iszero(x)
+        return zero(outT)
+    elseif isinf(x) && isinf(y)
+        return convert(outT,Inf)
+    else
+        return convert(outT,@fastmath exp2(convert(Float32, y) * fastlog2(convert(Float32, x))))
     end
 end
+
 @inline fastpow(x, y) = x^y

test/downstream/Project.toml

@@ -3,6 +3,7 @@ Calculus = "49dc2e85-a5d0-5ad3-a950-438e2897f1b9"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
 DiffEqBase = "2b5f629d-d688-5b77-993f-72d75c75574e"
 DiffEqCallbacks = "459566f4-90b8-5000-8ac3-15dfb0a30def"
+Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 Measurements = "eff96d63-e80a-5855-80a2-b1b0885c5ab7"
 ModelingToolkit = "961ee093-0014-501f-94e3-6117800e7a78"

test/downstream/enzyme.jl

@@ -0,0 +1,23 @@
+using Enzyme, EnzymeTestUtils
+using DiffEqBase: fastlog2, fastpow
+using Test
+
+@testset "Fast pow - Enzyme forward rule" begin
+    @testset for RT in (Duplicated, DuplicatedNoNeed),
+        Tx in (Const, Duplicated),
+        Ty in (Const, Duplicated)
+        x = 3.0
+        y = 2.0
+        test_forward(fastpow, RT, (x, Tx), (y, Ty), atol=0.005, rtol=0.005)
+    end
+end
+
+@testset "Fast pow - Enzyme reverse rule" begin
+    @testset for RT in (Active,),
+        Tx in (Active,),
+        Ty in (Active,)
+        x = 2.0
+        y = 3.0
+        test_reverse(fastpow, RT, (x, Tx), (y, Ty), atol=0.001, rtol=0.001)
+    end
+end

test/fastpow.jl

@@ -1,4 +1,4 @@
-using DiffEqBase: fastlog2, _exp2, fastpow
+using DiffEqBase: fastlog2, fastpow
 using Test
 
 @testset "Fast log2" begin
@@ -7,15 +7,9 @@ using Test
     end
 end
 
-@testset "Exp2" begin
-    for x in -100:0.01:3
-        @test exp2(x)≈_exp2(Float32(x)) atol=1e-6
-    end
-end
-
 @testset "Fast pow" begin
-    @test fastpow(1, 1) isa Float32
-    @test fastpow(1.0, 1.0) isa Float32
+    @test fastpow(1, 1) isa Float64
+    @test fastpow(1.0, 1.0) isa Float64
     errors = [abs(^(x, y) - fastpow(x, y)) for x in 0.001:0.001:1, y in 0.08:0.001:0.5]
     @test maximum(errors) < 1e-4
-end
+end