Skip to content

Specialize any, all, count and findall on the predicate, and infer any and all as Bool #191

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
maleadt merged 2 commits into from
Oct 29, 2026
Merged

Specialize any, all, count and findall on the predicate, and infer any and all as Bool #191

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c452152
Specialize `any`, `all`, `count` and `findall` on the predicate
AzamatB Oct 27, 2026
584ec2f
Infer `any` and `all` as `Bool` again
AzamatB Oct 28, 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
10 changes: 5 additions & 5 deletions src/findall.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -257,9 +257,9 @@ function findall(values::AbstractArray; items::AbstractArray=keys(values),
findall_run(identity, values, items, s, scratch_buffers(s.plan, workspace, values, items))
end

function findall(pred, v::AbstractArray; items::AbstractArray=keys(v),
function findall(pred::F, v::AbstractArray; items::AbstractArray=keys(v),
backend::Union{Nothing, Backend}=nothing, alg::Algorithm=Auto(),
workspace=nothing)
workspace=nothing) where {F}
check_bool_result(pred, v)
s = findall_setup(v, items, true, backend, alg)
findall_run(pred, v, items, s, scratch_buffers(s.plan, workspace, v, items))
Expand All @@ -268,8 +268,8 @@ end
plan(::typeof(findall), values::AbstractArray; items::AbstractArray=keys(values),
backend=nothing, alg::Algorithm=Auto()) =
findall_setup(values, items, false, backend, alg).plan
plan(::typeof(findall), pred, v::AbstractArray; items::AbstractArray=keys(v), backend=nothing,
alg::Algorithm=Auto()) =
plan(::typeof(findall), pred::F, v::AbstractArray; items::AbstractArray=keys(v),
backend=nothing, alg::Algorithm=Auto()) where {F} =
findall_setup(v, items, true, backend, alg).plan

# The plan of `findall` over `v`: its scratch is a mask of `pred(v[i])` for the predicate form
Expand All @@ -291,7 +291,7 @@ function findall_setup(v, items, predicate::Bool, backend, alg)
return (; plan=Plan(backend, a, merge(mask, counts)))
end

function findall_run(pred, v, items, s, bufs)
function findall_run(pred::F, v, items, s, bufs) where {F}
backend, a = s.plan.backend, s.plan.alg
bools = if haskey(bufs, :mask)
findall_mask!(bufs.mask, pred, v, backend, a)
Expand Down
34 changes: 18 additions & 16 deletions src/predicates.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,19 +162,20 @@ end
complex_any(CuArray(rand(Float32, 100)), CuArray(rand(Float32, 100)))
```
"""
function any(pred, v::AbstractArray; backend::Union{Nothing, Backend}=nothing,
alg::Algorithm=Auto(), workspace=nothing)
function any(pred::F, v::AbstractArray; backend::Union{Nothing, Backend}=nothing,
alg::Algorithm=Auto(), workspace=nothing) where {F}
p = predicate_setup(pred, |, false, v, backend, alg).plan
any_run(pred, v, p.backend, p.alg, scratch_buffers(p, workspace, v))
# (asserted, as `Auto()` picks a GPU's algorithm from its tuning, which inference does not see)
any_run(pred, v, p.backend, p.alg, scratch_buffers(p, workspace, v))::Bool
end

plan(::typeof(any), pred, v::AbstractArray; backend=nothing, alg::Algorithm=Auto()) =
plan(::typeof(any), pred::F, v::AbstractArray; backend=nothing, alg::Algorithm=Auto()) where {F} =
predicate_setup(pred, |, false, v, backend, alg).plan

# The predicate must return a `Bool`, as in Base. Where inference shows it cannot, the call fails
# on the host, before launching, with an `ArgumentError` instead of an exception from the kernel
# (which GPU backends report without Base's `TypeError`).
function check_bool_result(pred, v)
function check_bool_result(pred::F, v) where {F}
isempty(v) && return nothing # (Base never calls the predicate then)
R = Base.promote_op(pred, eltype(v))
R === Union{} || Bool <: R || throw(ArgumentError(
Expand All @@ -184,7 +185,7 @@ end

# The plan of `any`/`all` (whose `ViaReduce` reduces with `op` from `init`): the flag of
# `ConcurrentWrite`, or the scratch of the reduction
function predicate_setup(pred, op, init, v, backend, alg)
function predicate_setup(pred::F, op::OP, init, v, backend, alg) where {F, OP}
check_bool_result(pred, v)
backend = resolve_backend(backend, v)
a = resolve_predicate(alg, backend, eltype(v))
Expand All @@ -199,7 +200,7 @@ function predicate_setup(pred, op, init, v, backend, alg)
return (; plan=Plan(backend, a, sizes))
end

function any_run(pred, v, backend, alg::ConcurrentWrite, bufs)
function any_run(pred::F, v, backend, alg::ConcurrentWrite, bufs) where {F}
# `ndrange` must not be zero
isempty(v) && return false
out = bufs.flag
Expand All @@ -208,10 +209,10 @@ function any_run(pred, v, backend, alg::ConcurrentWrite, bufs)
return @allowscalar(out[1]) != 0
end

any_run(pred, v, backend, alg::ViaReduce, bufs) =
any_run(pred::F, v, backend, alg::ViaReduce, bufs) where {F} =
mapreduce_nested(BoolResult(pred), |, v, bufs.reduce; backend, init=false, alg=alg.reduce)

function any_run(pred, v, backend, alg::CPUThreads.Partitioned, bufs)
function any_run(pred::F, v, backend, alg::CPUThreads.Partitioned, bufs) where {F}
found = Threads.Atomic{Bool}(false)
first_index = firstindex(v)
task_partition(length(v), alg.max_tasks, alg.min_elems) do irange
Expand Down Expand Up @@ -252,17 +253,18 @@ end
complex_all(MtlArray(rand(Float32, 100)), MtlArray(rand(Float32, 100)))
```
"""
function all(pred, v::AbstractArray; backend::Union{Nothing, Backend}=nothing,
alg::Algorithm=Auto(), workspace=nothing)
function all(pred::F, v::AbstractArray; backend::Union{Nothing, Backend}=nothing,
alg::Algorithm=Auto(), workspace=nothing) where {F}
p = predicate_setup(pred, &, true, v, backend, alg).plan
all_run(pred, v, p.backend, p.alg, scratch_buffers(p, workspace, v))
all_run(pred, v, p.backend, p.alg, scratch_buffers(p, workspace, v))::Bool # (as in `any`)
end

plan(::typeof(all), pred, v::AbstractArray; backend=nothing, alg::Algorithm=Auto()) =
plan(::typeof(all), pred::F, v::AbstractArray; backend=nothing, alg::Algorithm=Auto()) where {F} =
predicate_setup(pred, &, true, v, backend, alg).plan

all_run(pred, v, backend, alg::ConcurrentWrite, bufs) = !any_run(!pred, v, backend, alg, bufs)
all_run(pred, v, backend, alg::ViaReduce, bufs) =
all_run(pred::F, v, backend, alg::ConcurrentWrite, bufs) where {F} =
!any_run(!pred, v, backend, alg, bufs)
all_run(pred::F, v, backend, alg::ViaReduce, bufs) where {F} =
mapreduce_nested(BoolResult(pred), &, v, bufs.reduce; backend, init=true, alg=alg.reduce)
all_run(pred, v, backend, alg::CPUThreads.Partitioned, bufs) =
all_run(pred::F, v, backend, alg::CPUThreads.Partitioned, bufs) where {F} =
!any_run(!pred, v, backend, alg, bufs)
4 changes: 2 additions & 2 deletions src/reduce/shorthands.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,13 +109,13 @@ c = AK.count(CuArray(rand(Bool, 10, 100_000)); init=Int32(0), dims=2)
```
"""
count(src::AbstractArray; kwargs...) = count(identity, src; kwargs...)
count(f, src::AbstractArray; init=0, kwargs...) =
count(f::F, src::AbstractArray; init=0, kwargs...) where {F} =
mapreduce(BoolResult(f), Base.add_sum, src; init, kwargs...)

plan(::typeof(sum), src::AbstractArray; kwargs...) = plan(reduce, Base.add_sum, src; kwargs...)
plan(::typeof(prod), src::AbstractArray; kwargs...) = plan(reduce, Base.mul_prod, src; kwargs...)
plan(::typeof(maximum), src::AbstractArray; kwargs...) = plan(reduce, max, src; kwargs...)
plan(::typeof(minimum), src::AbstractArray; kwargs...) = plan(reduce, min, src; kwargs...)
plan(::typeof(count), src::AbstractArray; kwargs...) = plan(count, identity, src; kwargs...)
plan(::typeof(count), f, src::AbstractArray; init=0, kwargs...) =
plan(::typeof(count), f::F, src::AbstractArray; init=0, kwargs...) where {F} =
plan(mapreduce, BoolResult(f), Base.add_sum, src; init, kwargs...)
31 changes: 31 additions & 0 deletions test/generic/predicates.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,10 @@ HOST_KERNELS || pushfirst!(PRED_ALGS, AK.Auto())
@test AK.all(x->x>0, v; alg) === true
@test AK.all(x->x<100, v; alg) === false

# The result infers, also where `Auto()` picks the algorithm from a GPU's tuning
@test @inferred(AK.any(x->x<0, v; alg)) === false
@test @inferred(AK.all(x->x>0, v; alg)) === true

for _ in 1:100
num_elems = rand(1:100_000)
v = array_from_host(rand(Float32, num_elems))
Expand Down Expand Up @@ -76,3 +80,30 @@ HOST_KERNELS || pushfirst!(PRED_ALGS, AK.Auto())
# Test that undefined kwargs are not accepted
@test_throws MethodError AK.any(x->x<0, v; alg, bad=:kwarg)
end

# Julia does not specialize a method on a function argument that it only passes on, so the
# methods that take the predicate name its type. Otherwise they run code compiled for any
# `Function`, in which `check_bool_result` infers the predicate's result type on every call.
# (Called by dynamic dispatch, as an inferred caller may inline past the entry point.)
@testset "predicates: no inference at run time" begin
# The compilation time of the fastest of a few warm calls (other tasks may compile once)
function compile_time(f, args...; kwargs...)
invokelatest(f, args...; kwargs...)
minimum(1:3) do _
Base.cumulative_compile_timing(true)
try
t = Base.cumulative_compile_time_ns()[1]
invokelatest(f, args...; kwargs...)
Base.cumulative_compile_time_ns()[1] - t
finally
Base.cumulative_compile_timing(false)
end
end
end
v = array_from_host(rand(Float32, 1000))
for alg in PRED_ALGS
@test compile_time(AK.any, isnan, v; alg) == 0
@test compile_time(AK.all, isfinite, v; alg) == 0
end
@test compile_time(AK.findall, isnan, v; alg=FINDALL_ALG) == 0
end
Loading