Skip to content

Commit

Permalink
make flattened Broadcasted more compiler friendly.
Browse files Browse the repository at this point in the history 
1. make `cat_nested` better inferred by switching to direct self-recursion.
2. `make_makeargs` now create a tuple of functions which take in the whole argument list and return the corresponding input for the broadcasted function.
  • Loading branch information
@N5N3
N5N3 committed Jul 15, 2023
1 parent 22ac24a commit d406c7e
Show file tree
Hide file tree
Showing 2 changed files with 59 additions and 81 deletions.
121 changes: 43 additions & 78 deletions base/broadcast.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -341,20 +341,16 @@ function flatten(bc::Broadcasted)
isflat(bc) && return bc
# concatenate the nested arguments into {a, b, c, d}
args = cat_nested(bc)
# build a function `makeargs` that takes a "flat" argument list and
# and creates the appropriate input arguments for `f`, e.g.,
# makeargs = (w, x, y, z) -> (w, g(x, y), z)
#
# `makeargs` is built recursively and looks a bit like this:
# makeargs(w, x, y, z) = (w, makeargs1(x, y, z)...)
# = (w, g(x, y), makeargs2(z)...)
# = (w, g(x, y), z)
let makeargs = make_makeargs(()->(), bc.args), f = bc.f
newf = @inline function(args::Vararg{Any,N}) where N
f(makeargs(args...)...)
end
return Broadcasted(bc.style, newf, args, bc.axes)
end
# build a tuple of functions `makeargs`. Its elements take
# the whole "flat" argument list and and generate the appropriate
# input arguments for the broadcasted function `f`, e.g.,
# makeargs[1] = ((w, x, y, z)) -> w
# makeargs[2] = ((w, x, y, z)) -> g(x, y)
# makeargs[3] = ((w, x, y, z)) -> z
makeargs = make_makeargs(bc.args)
f = Base.maybeconstructor(bc.f)
newf = (args...) -> (@inline; f(prepare_args(makeargs, args)...))
return Broadcasted(bc.style, newf, args, bc.axes)
end

const NestedTuple = Tuple{<:Broadcasted,Vararg{Any}}
Expand All @@ -363,78 +359,47 @@ _isflat(args::NestedTuple) = false
_isflat(args::Tuple) = _isflat(tail(args))
_isflat(args::Tuple{}) = true

cat_nested(t::Broadcasted, rest...) = (cat_nested(t.args...)..., cat_nested(rest...)...)
cat_nested(t::Any, rest...) = (t, cat_nested(rest...)...)
cat_nested() = ()
cat_nested(bc::Broadcasted) = cat_nested_args(bc.args)
cat_nested_args(::Tuple{}) = ()
cat_nested_args(t::Tuple{Any}) = cat_nested(t[1])
cat_nested_args(t::Tuple) = (cat_nested(t[1])..., cat_nested_args(tail(t))...)
cat_nested(a) = (a,)

"""
make_makeargs(makeargs_tail::Function, t::Tuple) -> Function
make_makeargs(t::Tuple) -> Tuple{Vararg{Function}}
Each element of `t` is one (consecutive) node in a broadcast tree.
Ignoring `makeargs_tail` for the moment, the job of `make_makeargs` is
to return a function that takes in flattened argument list and returns a
tuple (each entry corresponding to an entry in `t`, having evaluated
the corresponding element in the broadcast tree). As an additional
complication, the passed in tuple may be longer than the number of leaves
in the subtree described by `t`. The `makeargs_tail` function should
be called on such additional arguments (but not the arguments consumed
by `t`).
The returned `Tuple` are functions which take in the (whole) flattened
list and generate the inputs for the corresponding broadcasted function.
"""
@inline make_makeargs(makeargs_tail, t::Tuple{}) = makeargs_tail
@inline function make_makeargs(makeargs_tail, t::Tuple)
makeargs = make_makeargs(makeargs_tail, tail(t))
(head, tail...)->(head, makeargs(tail...)...)
make_makeargs(args::Tuple) = _make_makeargs(args, 1)[1]

# We build `makeargs` by traversing the broadcast nodes recursively.
# note: `n` indicates the flattened index of the next unused argument.
@inline function _make_makeargs(args::Tuple, n::Int)
head, n = _make_makeargs1(args[1], n)
rest, n = _make_makeargs(tail(args), n)
(head, rest...), n
end
function make_makeargs(makeargs_tail, t::Tuple{<:Broadcasted, Vararg{Any}})
bc = t[1]
# c.f. the same expression in the function on leaf nodes above. Here
# we recurse into siblings in the broadcast tree.
let makeargs_tail = make_makeargs(makeargs_tail, tail(t)),
# Here we recurse into children. It would be valid to pass in makeargs_tail
# here, and not use it below. However, in that case, our recursion is no
# longer purely structural because we're building up one argument (the closure)
# while destructuing another.
makeargs_head = make_makeargs((args...)->args, bc.args),
f = bc.f
# Create two functions, one that splits of the first length(bc.args)
# elements from the tuple and one that yields the remaining arguments.
# N.B. We can't call headargs on `args...` directly because
# args is flattened (i.e. our children have not been evaluated
# yet).
headargs, tailargs = make_headargs(bc.args), make_tailargs(bc.args)
return @inline function(args::Vararg{Any,N}) where N
args1 = makeargs_head(args...)
a, b = headargs(args1...), makeargs_tail(tailargs(args1...)...)
(f(a...), b...)
end
end
_make_makeargs(::Tuple{}, n::Int) = (), n

# A help struct to store the flattened index staticly
struct Pick{N} <: Function end
(::Pick{N})(@nospecialize(args::Tuple)) where {N} = args[N]

# For flat nodes, we just consume one argument (n += 1), and return the "Pick" function
@inline _make_makeargs1(_, n::Int) = Pick{n}(), n + 1
# For nested nodes, we form the `makeargs1` based on the child `makeargs` (n += length(cat_nested(bc)))
@inline function _make_makeargs1(bc::Broadcasted, n::Int)
makeargs, n = _make_makeargs(bc.args, n)
f = Base.maybeconstructor(bc.f)
makeargs1 = (args::Tuple) -> (@inline; f(prepare_args(makeargs, args)...))
makeargs1, n
end

@inline function make_headargs(t::Tuple)
let headargs = make_headargs(tail(t))
return @inline function(head, tail::Vararg{Any,N}) where N
(head, headargs(tail...)...)
end
end
end
@inline function make_headargs(::Tuple{})
return @inline function(tail::Vararg{Any,N}) where N
()
end
end

@inline function make_tailargs(t::Tuple)
let tailargs = make_tailargs(tail(t))
return @inline function(head, tail::Vararg{Any,N}) where N
tailargs(tail...)
end
end
end
@inline function make_tailargs(::Tuple{})
return @inline function(tail::Vararg{Any,N}) where N
tail
end
end
@inline prepare_args(makeargs::Tuple, @nospecialize(x::Tuple)) = (makeargs[1](x), prepare_args(tail(makeargs), x)...)
@inline prepare_args(makeargs::Tuple{Any}, @nospecialize(x::Tuple)) = (makeargs[1](x),)
prepare_args(::Tuple{}, ::Tuple) = ()

## Broadcasting utilities ##

Expand Down
19 changes: 16 additions & 3 deletions test/broadcast.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -774,14 +774,27 @@ let X = zeros(2, 3)
end

# issue #27988: inference of Broadcast.flatten
using .Broadcast: Broadcasted
using .Broadcast: Broadcasted, cat_nested
let
bc = Broadcasted(+, (Broadcasted(*, (1, 2)), Broadcasted(*, (Broadcasted(*, (3, 4)), 5))))
@test @inferred(Broadcast.cat_nested(bc)) == (1,2,3,4,5)
@test @inferred(cat_nested(bc)) == (1,2,3,4,5)
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == @inferred(Broadcast.materialize(bc)) == 62
bc = Broadcasted(+, (Broadcasted(*, (1, Broadcasted(/, (2.0, 2.5)))), Broadcasted(*, (Broadcasted(*, (3, 4)), 5))))
@test @inferred(Broadcast.cat_nested(bc)) == (1,2.0,2.5,3,4,5)
@test @inferred(cat_nested(bc)) == (1,2.0,2.5,3,4,5)
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == @inferred(Broadcast.materialize(bc)) == 60.8
# 1 .* 1 .- 1 .* 1 .^2 .+ 1 .* 1 .+ 1 .^ 3
bc = Broadcasted(+, (Broadcasted(+, (Broadcasted(-, (Broadcasted(*, (1, 1)), Broadcasted(*, (1, Broadcasted(Base.literal_pow, (Ref(^), 1, Ref(Val(2)))))))), Broadcasted(*, (1, 1)))), Broadcasted(Base.literal_pow, (Base.RefValue{typeof(^)}(^), 1, Base.RefValue{Val{3}}(Val{3}())))))
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == @inferred(Broadcast.materialize(bc)) == 2
# @. 1 + 1 * (1 + 1 + 1 + 1)
bc = Broadcasted(+, (1, Broadcasted(*, (1, Broadcasted(+, (1, 1, 1, 1))))))
@test @inferred(cat_nested(bc)) == (1, 1, 1, 1, 1, 1) # `cat_nested` failed to infer this
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == Broadcast.materialize(bc)
# @. 1 + (1 + 1) + 1 + (1 + 1) + 1 + (1 + 1) + 1
bc = Broadcasted(+, (1, Broadcasted(+, (1, 1)), 1, Broadcasted(+, (1, 1)), 1, Broadcasted(+, (1, 1)), 1))
@test @inferred(cat_nested(bc)) == (1, 1, 1, 1, 1, 1, 1, 1, 1, 1)
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == Broadcast.materialize(bc)
bc = Broadcasted(Float32, (Broadcasted(+, (1, 1)),))
@test @inferred(Broadcast.materialize(Broadcast.flatten(bc))) == Broadcast.materialize(bc)
end

let
Expand Down

0 comments on commit d406c7e

Please sign in to comment.