Skip to content

Specialize mapreduce glue to avoid boxing kernel arguments #87

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lkdvos merged 2 commits into from
Oct 7, 2026
Merged

Specialize mapreduce glue to avoid boxing kernel arguments #87

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
90b257e
specialize glue for the kernel mapping to avoid boxing variables
lkdvos Oct 7, 2026
027d5ff
add explanation comment
lkdvos Oct 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 21 additions & 20 deletions src/mapreduce.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,17 +40,22 @@ function Base._mapreduce_dim(f, op, ::NamedTuple{()}, A::StridedView, dims)
end

function Base.map(
@nospecialize(f), a1::StridedView{<:Any, N},
f::F, a1::StridedView{<:Any, N},
A::Vararg{StridedView{<:Any, N}}
) where {N}
) where {F, N}
T = Base.promote_op(f, eltype(a1), eltype.(A)...)
return map!(f, similar(a1, T), a1, A...)
end

# Note that here and below we have to specialize on the `f` and `op` arguments.
# This is because the @generated kernel always specializes, so if the callers don't
# we end up with a dynamic dispatch that has to box these variables, causing allocations
# and overhead

function Base.map!(
@nospecialize(f), b::StridedView{<:Any, N}, a1::StridedView{<:Any, N},
f::F, b::StridedView{<:Any, N}, a1::StridedView{<:Any, N},
A::Vararg{StridedView{<:Any, N}}
) where {N}
) where {F, N}
dims = size(b)

# Check dimesions
Expand All @@ -66,7 +71,7 @@ function Base.map!(
return b
end

function _mapreduce(@nospecialize(f), @nospecialize(op), A::StridedView{T}, nt = nothing) where {T}
function _mapreduce(f::F, op::OP, A::StridedView{T}, nt = nothing) where {F, OP, T}
if isempty(A)
b = Base.mapreduce_empty(f, op, T)
return nt === nothing ? b : op(b, nt.init)
Expand All @@ -87,10 +92,10 @@ function _mapreduce(@nospecialize(f), @nospecialize(op), A::StridedView{T}, nt =
end

function Base.mapreducedim!(
@nospecialize(f), @nospecialize(op), b::StridedView{<:Any, N},
f::F, op::OP, b::StridedView{<:Any, N},
a1::StridedView{<:Any, N},
A::Vararg{StridedView{<:Any, N}}
) where {N}
) where {F, OP, N}
outdims = size(b)
dims = map(max, outdims, map(max, map(size, (a1, A...))...))

Expand All @@ -101,9 +106,8 @@ function Base.mapreducedim!(
end

function _mapreducedim!(
@nospecialize(f), @nospecialize(op), @nospecialize(initop),
dims::Dims, arrays::Tuple{Vararg{StridedView}}
)
f::F, op::OP, initop::I, dims::Dims, arrays::Tuple{Vararg{StridedView}}
) where {F, OP, I}
if any(isequal(0), dims)
if length(arrays[1]) != 0 && !isnothing(initop)
map!(initop, arrays[1], arrays[1])
Expand Down Expand Up @@ -167,9 +171,8 @@ _computecosts(strides) = map(a -> ifelse(iszero(a), 1, a << 1), map(min, strides
# Pipeline entry point: order → fuse → order → block → kernel.
# Fusing needs ordered entries, and final order pass brings remaining dim 1 to end
function _mapreduce_order!(
@nospecialize(f), @nospecialize(op), @nospecialize(initop),
dims::Dims, arrays::Tuple{Vararg{StridedView}}
)
f::F, op::OP, initop::I, dims::Dims, arrays::Tuple{Vararg{StridedView}}
) where {F, OP, I}
isempty(dims) && return _mapreduce_scalar!(f, op, initop, arrays)
dims, allstrides = order_and_fuse_dims(dims, map(strides, arrays))
offsets = map(offset, arrays)
Expand All @@ -178,7 +181,7 @@ function _mapreduce_order!(
end

# 0-dimensional fast path: bypass @generated kernel
function _mapreduce_scalar!(@nospecialize(f), @nospecialize(op), @nospecialize(initop), arrays)
function _mapreduce_scalar!(f::F, op::OP, initop::I, arrays) where {F, OP, I}
out = arrays[1]
iout = ParentIndex(offset(out) + 1)
v = f(map(a -> a[ParentIndex(offset(a) + 1)], Base.tail(arrays))...)
Expand All @@ -193,9 +196,8 @@ end

const MINTHREADLENGTH = 1 << 15 # minimal length before any kind of threading is applied
function _mapreduce_block!(
@nospecialize(f), @nospecialize(op), @nospecialize(initop),
dims, strides, offsets, costs, arrays
)
f::F, op::OP, initop::I, dims, strides, offsets, costs, arrays
) where {F, OP, I}
bytestrides = map((s, stride) -> s .* stride, sizeof.(eltype.(arrays)), strides)
strideorders = map(indexorder, strides)
blocks = _computeblocks(dims, costs, bytestrides, strideorders)
Expand Down Expand Up @@ -254,10 +256,9 @@ end
# nthreads: number of threads spacing: extra addition to offset of array 1, to account for
# reduction
function _mapreduce_threaded!(
@nospecialize(f), @nospecialize(op), @nospecialize(initop),
dims, blocks, strides, offsets, costs, arrays, nthreads,
f::F, op::OP, initop::I, dims, blocks, strides, offsets, costs, arrays, nthreads,
spacing, taskindex
)
) where {F, OP, I}
if nthreads == 1 || prod(dims) <= MINTHREADLENGTH
offset1 = offsets[1] + spacing * (taskindex - 1)
spacedoffsets = (offset1, Base.tail(offsets)...)
Expand Down
Loading