Unexplained allocations

[shashi]

(in #3)

A rather scathing benchmark:

julia> using ForwardDiff2: dualrun

julia> @btime rand()
  3.199 ns (0 allocations: 0 bytes)
0.8383875026881109

julia> @btime dualrun(()->rand())
  930.333 ns (5 allocations: 112 bytes)
0.7136361983030461

So I tried to track allocations for

julia> foo() = for i=1:10^6; dualrun(()->rand()); end

Which gave me:

        - @inline function overdub(ctx::TaggedCtx{T}, f, args...) where {T}
287281248     if length(args) > 4
        -         return Cassette.recurse(ctx, f, args...)
        -     end
        -     # find the position of the dual number with the highest
        -     # precedence (dominant) tag
 23520864     idx = find_dual(Tag{T}, args...)
        0     if idx === 0
        -         # none of the arguments are dual
 70562592         Cassette.recurse(ctx, f, args...)
        -     else

The file containing find_dual itself shows no allocations. There are in fact no allocations anywhere else within ForwardDiff2.

@oxinabox any help?

[oxinabox]

I am on holidays. Poke me in a 3+ days

[ChrisRackauckas]

3+ days @oxinabox 😆

[oxinabox]

@inline function overdub(ctx::TaggedCtx{T}, f, arg1, arg2, arg3, arg4, arg5, args...) where {T}
     return Cassette.recurse(ctx, args)
end

@inline function overdub(ctx::TaggedCtx{T}, f, args) where {T}
    # find the position of the dual number with the highest
	# precedence (dominant) tag
	idx = find_dual(Tag{T}, args...)
	if idx === 0
		# none of the arguments are dual
		Cassette.recurse(ctx, f, args...)
	else
	    ...
	end
end

Or (since recursing is the default):

for nargs in 1:4
    arg_names = ntuple(n->Symbol(:arg, n), nargs)
	
	@eval @inline function overdub(ctx::TaggedCtx{T}, f, $(arg_names...)) where {T}
		# find the position of the dual number with the highest
		# precedence (dominant) tag
		idx = find_dual(Tag{T}, $(arg_names...))
		if idx === 0
			# none of the arguments are dual
			Cassette.recurse(ctx, f, $(arg_names...))
		else
			...
		end
	end
end

Which will remove the runtime splatting entirely

[ChrisRackauckas]

@YingboMa was this solved?

[YingboMa]

It hasn't been solved yet, but there is an improvement.

julia> using ForwardDiff2: dualrun

julia> using BenchmarkTools

julia> @btime dualrun(()->rand())
  453.505 ns (1 allocation: 32 bytes)
0.6022948398960193