Skip to content

Add native BFloat16 atomic_add! support #3007

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

Draft
AntonOresten wants to merge 2 commits into
base: master
Choose a base branch
from
+13 −8
Draft

Add native BFloat16 atomic_add! support #3007

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
3a9cda6
Bind PTX for atomic add of BFloat16
AntonOresten Jan 2, 2026
9672643
fix CC requirement
AntonOresten Jan 2, 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
9 changes: 5 additions & 4 deletions src/device/intrinsics/atomics.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,18 +180,19 @@ end

## PTX

# half-precision atomics using PTX instruction
# half-precision and bfloat16 atomics using PTX instructions

for A in (AS.Generic, AS.Global, AS.Shared), T in (:Float16,)
for A in (AS.Generic, AS.Global, AS.Shared), T in (:Float16, :BFloat16)
if A == AS.Global
scope = ".global"
elseif A == AS.Shared
scope = ".shared"
else
scope = ""
end
name = T == :Float16 ? "f16" : "bf16"

intr = "atom$scope.add.noftz.f16 \$0, [\$1], \$2;"
intr = "atom$scope.add.noftz.$name \$0, [\$1], \$2;"
@eval @inline atomic_add!(ptr::LLVMPtr{$T,$A}, val::$T) =
@asmcall($intr, "=h,l,h", true, $T, Tuple{Core.LLVMPtr{$T,$A},$T}, ptr, val)
end
Expand Down Expand Up @@ -441,7 +442,7 @@ end
atomic_arrayset(A, Base._to_linear_index(A, Is...), op, convert(T, val))

# native atomics
for (op,impl,typ) in [(:(+), :(atomic_add!), [:UInt32,:Int32,:UInt64,:Int64,:Float32]),
for (op,impl,typ) in [(:(+), :(atomic_add!), [:UInt32,:Int32,:UInt64,:Int64,:Float32,:Float16,:BFloat16]),
(:(-), :(atomic_sub!), [:UInt32,:Int32,:UInt64,:Int64,:Float32]),
(:(&), :(atomic_and!), [:UInt32,:Int32,:UInt64,:Int64]),
(:(|), :(atomic_or!), [:UInt32,:Int32,:UInt64,:Int64]),
Expand Down
12 changes: 8 additions & 4 deletions test/core/device/intrinsics/atomics.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@ using BFloat16s: BFloat16
types = [Int32, Int64, UInt32, UInt64, Float32]
capability(device()) >= v"6.0" && push!(types, Float64)
capability(device()) >= v"7.0" && push!(types, Float16)
capability(device()) >= v"9.0" && push!(types, BFloat16)

@testset for T in types
a = CuArray(T[0])
Expand All @@ -19,8 +20,9 @@ using BFloat16s: BFloat16
return
end

@cuda threads=1024 kernel(a, one(T))
@test Array(a)[1] == 1024
nthreads = T == BFloat16 ? 128 : 1024 # BFloat16(256) + 1 == 256
@cuda threads=nthreads kernel(a, one(T))
@test Array(a)[1] == nthreads
end
end

Expand Down Expand Up @@ -212,6 +214,7 @@ end
@testset "add" begin
types = [Int32, Int64, UInt32, UInt64, Float32, Float64]
capability(device()) >= v"7.0" && append!(types, [Int16, UInt16, Float16])
capability(device()) >= v"9.0" && push!(types, BFloat16)

@testset for T in types
a = CuArray([zero(T)])
Expand All @@ -222,8 +225,9 @@ end
return
end

@cuda threads=1024 kernel(T, a)
@test Array(a)[1] == 2048
nthreads = T == BFloat16 ? 64 : 1024
@cuda threads=nthreads kernel(T, a)
@test Array(a)[1] == 2 * nthreads
end
end

Expand Down