Skip to content
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

Introduce AsyncNumber to lazily copy numeric mapreduce results to the host #550

Closed
wants to merge 8 commits into from
Closed

Introduce AsyncNumber to lazily copy numeric mapreduce results to the host #550

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
cf88dfe
Add GPUNumber
pxl-th Jul 6, 2024
16bb79f
Add compat
pxl-th Jul 6, 2024
aa095e8
Add more methods
pxl-th Jul 7, 2024
982e7cc
Add more methods
pxl-th Jul 7, 2024
59f0a88
More methods...
pxl-th Jul 7, 2024
8e48354
Return GPUNumber only when eltype is Number
pxl-th Jul 9, 2024
d951132
Cleanup
pxl-th Jul 9, 2024
f01a7d1
Refactoring
pxl-th Jul 10, 2024
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
2 changes: 2 additions & 0 deletions Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@ uuid = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
version = "10.2.3"

[deps]
AbstractNumbers = "85c772de-338a-5e7f-b815-41e76c26ac1f"
Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
Expand All @@ -14,6 +15,7 @@ Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"

[compat]
AbstractNumbers = "0.2"
Adapt = "4.0"
GPUArraysCore = "= 0.1.6"
LLVM = "3.9, 4, 5, 6, 7, 8"
Expand Down
3 changes: 3 additions & 0 deletions src/GPUArrays.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
module GPUArrays

import AbstractNumbers as AN

using Serialization
using Random
using LinearAlgebra
Expand All @@ -25,6 +27,7 @@ include("device/synchronization.jl")
# host abstractions
include("host/abstractarray.jl")
include("host/construction.jl")
include("host/gpunumber.jl")
## integrations and specialized methods
include("host/base.jl")
include("host/indexing.jl")
Expand Down
42 changes: 42 additions & 0 deletions src/host/gpunumber.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
# Custom GPU-compatible `Number` interface.
struct GPUNumber{T <: AbstractGPUArray} <: AN.AbstractNumber{T}
pxl-th marked this conversation as resolved.
Show resolved Hide resolved
val::T

function GPUNumber(val::T) where T <: AbstractGPUArray
Copy link
Contributor

@vpuri3 vpuri3 Jul 8, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@pxl-th it might be a good idea to do a dropdims or vec in the constructor so this guy truly behaves like a number. The case I am thinking of is the following.

a = CUDA.rand(1,1,1,1) |> GPUNumber
x = CUDA.rand(4,4)
size(x .+ a) # want (4,4) but will likely get (4,4,1,1)

I haven't run your code but it seems like it would do something similar to

julia> rand(1,1,1,1) .+ rand(4,4) |> size
(4, 4, 1, 1)

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Indeed, thanks!

length(val) != 1 && error(
"`GPUNumber` accepts only 1-element GPU arrays, " *
"instead `$(length(val))`-element array was given.")
new{T}(val)
end
end

AN.number(g::GPUNumber) = @allowscalar g.val[]
maybe_number(g::GPUNumber) = AN.number(g)
maybe_number(g) = g

number_type(::GPUNumber{T}) where T = eltype(T)

# When operations involve other `::Number` types,
# do not convert back to `GPUNumber`.
AN.like(::Type{<: GPUNumber}, x) = x

# When broadcasting, just pass the array itself.
Base.broadcastable(g::GPUNumber) = g.val

# Overload to avoid copies.
Base.one(g::GPUNumber) = one(number_type(g))
Base.one(::Type{GPUNumber{T}}) where T = one(eltype(T))
Base.zero(g::GPUNumber) = zero(number_type(g))
Base.zero(::Type{GPUNumber{T}}) where T = zero(eltype(T))
Base.identity(g::GPUNumber) = g

Base.getindex(g::GPUNumber) = AN.number(g)

Base.isequal(g::GPUNumber, v::Number) = isequal(g[], v)
Base.isequal(v::Number, g::GPUNumber) = isequal(v, g[])

Base.nextpow(a, x::GPUNumber) = nextpow(a, x[])
Base.nextpow(a::GPUNumber, x) = nextpow(a[], x)
Base.nextpow(a::GPUNumber, x::GPUNumber) = nextpow(a[], x[])

Base.convert(::Type{Number}, g::GPUNumber) = g[]
6 changes: 3 additions & 3 deletions src/host/indexing.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,8 +203,7 @@ function Base.findfirst(f::Function, A::AnyGPUArray)
return (false, dummy_index)
end

res = mapreduce((x, y)->(f(x), y), reduction, A, indices;
init = (false, dummy_index))
res = mapreduce((x, y)->(f(x), y), reduction, A, indices; init = (false, dummy_index))
if res[1]
# out of consistency with Base.findarray, return a CartesianIndex
# when the input is a multidimensional array
Expand All @@ -230,7 +229,8 @@ function findminmax(binop, A::AnyGPUArray; init, dims)
end

if dims == Colon()
res = mapreduce(tuple, reduction, A, indices; init = (init, dummy_index))
res = mapreduce(tuple, reduction, A, indices;
init = (init, dummy_index))

# out of consistency with Base.findarray, return a CartesianIndex
# when the input is a multidimensional array
Expand Down
22 changes: 13 additions & 9 deletions src/host/mapreduce.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,20 +68,23 @@ function _mapreduce(f::F, op::OP, As::Vararg{Any,N}; dims::D, init) where {F,OP,
end

if dims === Colon()
@allowscalar R[]
# Return `GPUNumber` for `Number` eltypes, otherwise - transfer to host.
eltype(R) <: Number ?
GPUNumber(reshape(R, :)) :
@allowscalar(R[])
else
R
end
end

Base.any(A::AnyGPUArray{Bool}) = mapreduce(identity, |, A)
Base.all(A::AnyGPUArray{Bool}) = mapreduce(identity, &, A)
Base.any(A::AnyGPUArray{Bool}) = mapreduce(identity, |, A)[]
Base.all(A::AnyGPUArray{Bool}) = mapreduce(identity, &, A)[]

Base.any(f::Function, A::AnyGPUArray) = mapreduce(f, |, A)
Base.all(f::Function, A::AnyGPUArray) = mapreduce(f, &, A)
Base.any(f::Function, A::AnyGPUArray) = mapreduce(f, |, A)[]
Base.all(f::Function, A::AnyGPUArray) = mapreduce(f, &, A)[]

Base.count(pred::Function, A::AnyGPUArray; dims=:, init=0) =
mapreduce(pred, Base.add_sum, A; init=init, dims=dims)
mapreduce(pred, Base.add_sum, A; init=init, dims=dims) |> maybe_number

# avoid calling into `initarray!`
for (fname, op) in [(:sum, :(Base.add_sum)), (:prod, :(Base.mul_prod)),
Expand All @@ -94,7 +97,7 @@ for (fname, op) in [(:sum, :(Base.add_sum)), (:prod, :(Base.mul_prod)),
end
end

LinearAlgebra.ishermitian(A::AbstractGPUMatrix) = mapreduce(==, &, A, adjoint(A))
LinearAlgebra.ishermitian(A::AbstractGPUMatrix) = mapreduce(==, &, A, adjoint(A))[]


# comparisons
Expand All @@ -105,7 +108,7 @@ function Base.isequal(A::AnyGPUArray, B::AnyGPUArray)
if axes(A) != axes(B)
return false
end
mapreduce(isequal, &, A, B; init=true)
mapreduce(isequal, &, A, B; init=true)[]
end

# returns `missing` when missing values are involved
Expand All @@ -129,6 +132,7 @@ function Base.:(==)(A::AnyGPUArray, B::AnyGPUArray)
(; is_missing=false, is_equal=a.is_equal & b.is_equal)
end
end
res = mapreduce(mapper, reducer, A, B; init=(; is_missing=false, is_equal=true))
res = mapreduce(mapper, reducer, A, B;
init=(; is_missing=false, is_equal=true))
res.is_missing ? missing : res.is_equal
end
Loading