fixup, rm many comments

Author: mcabbott

src/rulesets/Base/base.jl

@@ -249,13 +249,17 @@ end
 #####
 
 function rrule(cfg::RuleConfig{>:HasReverseMode}, ::typeof(map), f::F, x::AbstractArray) where {F}
-    # y, back = rrule_via_ad(cfg, Broadcast.broadcasted, f, x) # Yota likes this one
-    # return Broadcast.materialize(y), back
-    y, back = rrule_via_ad(cfg, broadcast, f, x)  # but testing like this one
+    # Here map agrees with broadcast, and we have a meta-rule with 4 different paths, should be fast:
+    y, back = rrule_via_ad(cfg, broadcast, f, x)
     return y, back
 end
 
 function rrule(cfg::RuleConfig{>:HasReverseMode}, ::typeof(map), f::F, x::AbstractArray, ys::AbstractArray...) where {F}
+    if all(==(size(x)), map(size, ys))
+        # Here too map agrees with broadcast, maybe the test could be more elegant?
+        y, back = rrule_via_ad(cfg, broadcast, f, x, ys...)
+        return y, back
+    end
     @debug "rrule(map, f, arrays...)" f
     z, backs = unzip_map((xy...) -> rrule_via_ad(cfg, f, xy...), x, ys...)
     function map_pullback_2(dz)

src/rulesets/Base/broadcast.jl

@@ -121,26 +121,6 @@ end
 # Path 4: The most generic, save all the pullbacks. Can be 1000x slower.
 # While broadcast makes no guarantee about order of calls, it's cheap to reverse the iteration.
 
-#=
-
-julia> Yota.grad(xs -> sum(abs2, (x -> abs(x)).(xs)), [1,2,3.0])
-┌ Debug: split broadcasting generic
-│   f = #69 (generic function with 1 method)
-│   N = 1
-└ @ ChainRules ~/.julia/dev/ChainRules/src/rulesets/Base/broadcast.jl:126
-(14.0, (ZeroTangent(), [2.0, 4.0, 6.0]))
-
-julia> ENV["JULIA_DEBUG"] = nothing
-
-julia> @btime Yota.grad(xs -> sum(abs2, (x -> abs(x)).(xs)), $(rand(1000)));
-  min 1.321 ms, mean 1.434 ms (23010 allocations, 594.66 KiB)  # with unzip_map, as before
-  min 1.279 ms, mean 1.393 ms (23029 allocations, 595.73 KiB)  # with unzip_map_reversed
-
-julia> @btime Yota.grad(xs -> sum(abs2, abs.(xs)), $(randn(1000)));  # Debug: split broadcasting derivative
-  min 2.144 μs, mean 6.620 μs (6 allocations, 23.88 KiB)
-
-=#
-
 function split_bc_pullbacks(cfg::RCR, f::F, args::Vararg{Any,N}) where {F,N}
     @debug("split broadcasting generic", f, N)
     ys3, backs = unzip_broadcast(args...) do a...

src/rulesets/Base/iterators.jl

@@ -17,57 +17,10 @@ function rrule(cfg::RuleConfig{>:HasReverseMode}, ::typeof(collect), gen::G) whe
     ys, generator_pullback
 end
 
-# Needed for Yota, but shouldn't these be automatic?
-ChainRulesCore.rrule(::Type{<:Base.Generator}, f, iter) = Base.Generator(f, iter), dy -> (NoTangent(), dy.f, dy.iter)
-ChainRulesCore.rrule(::Type{<:Iterators.ProductIterator}, iters) = Iterators.ProductIterator(iters), dy -> (NoTangent(), dy.iterators)
-
-#=
-
-          Yota.grad(xs -> sum(abs, [sin(x) for x in xs]), [1,2,3]pi/3)
-Diffractor.gradient(xs -> sum(abs, [sin(x) for x in xs]), [1,2,3]pi/3)
-
-          Yota.grad((xs, ys) -> sum(abs, [atan(x/y) for x in xs, y in ys]), [1,2,3]pi/3, [4,5])  # ERROR: all field arrays must have same shape
-Diffractor.gradient((xs, ys) -> sum(abs, [atan(x/y) for x in xs, y in ys]), [1,2,3]pi/3, [4,5])  # ERROR: type Array has no field iterators
-
-          Yota.grad(xs -> sum(abs, map(sin, xs)), [1,2,3]pi/3)
-Diffractor.gradient(xs -> sum(abs, map(sin, xs)), [1,2,3]pi/3)  # fails internally
-
-          Yota.grad(xs -> sum(abs, [sin(x/y) for (x,y) in zip(xs, 1:2)]), [1,2,3]pi/3)
-Diffractor.gradient(xs -> sum(abs, [sin(x/y) for (x,y) in zip(xs, 1:2)]), [1,2,3]pi/3)
-
-          Yota.grad(xs -> sum(abs, map((x,y) -> sin(x/y), xs, 1:2)), [1,2,3]pi/3)
-Diffractor.gradient(xs -> sum(abs, map((x,y) -> sin(x/y), xs, 1:2)), [1,2,3]pi/3)
-
-
-@btime Yota.grad($(rand(1000))) do xs
-    sum(abs2, [sqrt(x) for x in xs])
-end
-# Yota         min 759.000 μs, mean 800.754 μs (22041 allocations, 549.62 KiB)
-# Diffractor   min 559.000 μs, mean 622.464 μs (18051 allocations, 612.34 KiB)
-
-# Zygote  min 3.198 μs, mean 6.849 μs (20 allocations, 40.11 KiB)
-
-
-@btime Yota.grad($(rand(1000)), $(rand(1000))) do xs, ys
-    zs = map(xs, ys) do x, y
-        atan(x/y)
-    end
-    sum(abs2, zs)
-end
-# Yota + CR:      min 1.598 ms, mean 1.691 ms (38030 allocations, 978.75 KiB)
-# Diffractor + CR:  min 767.250 μs, mean 847.640 μs (26045 allocations, 838.66 KiB)
-
-# Zygote: min 13.417 μs, mean 22.896 μs (26 allocations, 79.59 KiB) -- 100x faster
-
-
-=#
-
-
 #####
 ##### `zip`
 #####
 
-
 function rrule(::typeof(zip), xs::AbstractArray...)
     function zip_pullback(dy)
         @debug "zip array pullback" summary(dy)
@@ -94,8 +47,6 @@ function _unmap_pad(x::AbstractArray, dx::AbstractArray)
         @debug "_unmap_pad is extending gradient" length(x) == length(dx)
         i1 = firstindex(x)
         ∇getindex(x, vec(dx), i1:i1+length(dx)-1)
-        # dx2 = vcat(vec(dx), similar(x, ZeroTangent, length(x) - length(dx)))
-        # ProjectTo(x)(reshape(dx2, axes(x)))
     end 
 end
 

test/rulesets/Base/base.jl

@@ -251,6 +251,8 @@
 
     @testset "map(f, ::Array, ::Array)" begin
         test_rrule(map, +, [1.0, 2.0], [3.0, 4.0], check_inferred=false)  # NoTangent does not match Union{NoTangent, ZeroTangent}
+        test_rrule(map, atan, [1 2; 3.0 4.0], [4 5; 6 7.0], check_inferred=false)  # same shape => just broadcast
+
         test_rrule(map, /, [1.0, 2.0], [3.0, 4.0, 5.0], check_inferred=false)
         test_rrule(map, atan, [1, 2, 3.0], [4 5; 6 7.0], check_inferred=false)
 

test/rulesets/Base/iterators.jl

@@ -10,7 +10,7 @@
 
         y2, bk2 = rrule(CFG, collect, Iterators.map(Counter(), [11, 12, 13.0]))
         @test y2 == map(Counter(), 11:13)
-        @test bk2(ones(3))[2].iter == [93, 83, 73]
+        @test bk2(ones(3))[2].iter == [33, 23, 13]
     end
 end
 
@@ -23,4 +23,4 @@ end
         test_rrule(collect∘zip, rand(3), rand(5))
         test_rrule(collect∘zip, rand(3,2), rand(5))
     end
-end
+end