Merge pull request #492 from Rabab53/matmul

DArray: Implement efficient in-place matrix-matrix multiply

Author: jpsamaroo

.buildkite/pipeline.yml

@@ -4,6 +4,7 @@
     queue: "juliaecosystem"
     sandbox_capable: "true"
     os: linux
+    arch: x86_64
   command: "julia --project -e 'using Pkg; Pkg.develop(;path=\"lib/TimespanLogging\")'"
 .bench: &bench
   if: build.message =~ /\[run benchmarks\]/
@@ -15,7 +16,7 @@
     num_cpus: 16
 steps:
   - label: Julia 1.8
-    timeout_in_minutes: 60
+    timeout_in_minutes: 90
     <<: *test
     plugins:
       - JuliaCI/julia#v1:
@@ -25,7 +26,7 @@ steps:
       - JuliaCI/julia-coverage#v1:
           codecov: true
   - label: Julia 1.9
-    timeout_in_minutes: 60
+    timeout_in_minutes: 90
     <<: *test
     plugins:
       - JuliaCI/julia#v1:
@@ -35,7 +36,7 @@ steps:
       - JuliaCI/julia-coverage#v1:
           codecov: true
   - label: Julia 1.10
-    timeout_in_minutes: 60
+    timeout_in_minutes: 90
     <<: *test
     plugins:
       - JuliaCI/julia#v1:
@@ -45,7 +46,7 @@ steps:
       - JuliaCI/julia-coverage#v1:
           codecov: true
   - label: Julia nightly
-    timeout_in_minutes: 60
+    timeout_in_minutes: 90
     <<: *test
     plugins:
       - JuliaCI/julia#v1:
@@ -55,7 +56,7 @@ steps:
       - JuliaCI/julia-coverage#v1:
           codecov: true
   - label: Julia 1.8 (macOS)
-    timeout_in_minutes: 60
+    timeout_in_minutes: 90
     <<: *test
     agents:
       queue: "juliaecosystem"
@@ -69,7 +70,7 @@ steps:
       - JuliaCI/julia-coverage#v1:
           codecov: true
   - label: Julia 1.8 - TimespanLogging
-    timeout_in_minutes: 60
+    timeout_in_minutes: 20
     <<: *test
     plugins:
       - JuliaCI/julia#v1:
@@ -78,7 +79,7 @@ steps:
           codecov: true
     command: "julia --project -e 'using Pkg; Pkg.instantiate(); Pkg.develop(;path=\"lib/TimespanLogging\"); Pkg.test(\"TimespanLogging\")'"
   - label: Julia 1.8 - DaggerWebDash
-    timeout_in_minutes: 60
+    timeout_in_minutes: 20
     <<: *test
     plugins:
       - JuliaCI/julia#v1:

Project.toml

@@ -20,6 +20,7 @@ SharedArrays = "1a1011a3-84de-559e-8e89-a11a2f7dc383"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
+TaskLocalValues = "ed4db957-447d-4319-bfb6-7fa9ae7ecf34"
 TimespanLogging = "a526e669-04d3-4846-9525-c66122c55f63"
 UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 
@@ -34,6 +35,7 @@ Requires = "1"
 ScopedValues = "1.1"
 Statistics = "1"
 StatsBase = "0.28, 0.29, 0.30, 0.31, 0.32, 0.33, 0.34"
+TaskLocalValues = "0.1"
 TimespanLogging = "0.1"
 julia = "1.8"
 

docs/src/darray.md

@@ -134,12 +134,7 @@ julia> DZ = DY .* 3
 Dagger.DArray{Float64, 2, Blocks{2}, typeof(cat)}(100, 100)
 ```
 
-Now, `DZ` will contain the result of computing `(DX .+ DX) .* 3`. Note that
-`DArray` objects are immutable, and operations on them are thus functional
-transformations of their input `DArray`.
-
-!!! note
-    Support for mutation of `DArray`s is planned for a future release
+Now, `DZ` will contain the result of computing `(DX .+ DX) .* 3`.
 
 ```
 julia> Dagger.chunks(DZ)
@@ -208,15 +203,17 @@ julia> collect(DZ)
 ```
 
 A variety of other operations exist on the `DArray`, and it should generally
-behavior otherwise similar to any other `AbstractArray` type. If you find that
+behave otherwise similar to any other `AbstractArray` type. If you find that
 it's missing an operation that you need, please file an issue!
 
 ### Known Supported Operations
 
 This list is not exhaustive, but documents operations which are known to work well with the `DArray`:
 
 From `Base`:
+- `getindex`/`setindex!`
 - Broadcasting
+- `similar`/`copy`/`copyto!`
 - `map`/`reduce`/`mapreduce`
 - `sum`/`prod`
 - `minimum`/`maximum`/`extrema`
@@ -225,3 +222,9 @@ From `Statistics`:
 - `mean`
 - `var`
 - `std`
+
+From `LinearAlgebra`:
+- `transpose`/`adjoint` (Out-of-place transpose)
+- `*` (Out-of-place Matrix-(Matrix/Vector) multiply)
+- `mul!` (In-place Matrix-Matrix multiply)
+- `cholesky`/`cholesky!` (In-place/Out-of-place Cholesky factorization)

src/Dagger.jl

@@ -60,17 +60,19 @@ include("datadeps.jl")
 include("array/darray.jl")
 include("array/alloc.jl")
 include("array/map-reduce.jl")
+include("array/copy.jl")
 
 # File IO
 include("file-io.jl")
 
 include("array/operators.jl")
-include("array/getindex.jl")
+include("array/indexing.jl")
 include("array/setindex.jl")
 include("array/matrix.jl")
 include("array/sparse_partition.jl")
 include("array/sort.jl")
 include("array/linalg.jl")
+include("array/mul.jl")
 include("array/cholesky.jl")
 
 # Visualization

src/array/copy.jl

@@ -0,0 +1,111 @@
+# Copy Buffering
+
+function maybe_copy_buffered(f, args...)
+    @assert all(arg->arg isa Pair{<:DArray,<:Blocks}, args) "maybe_copy_buffered only supports `DArray`=>`Blocks`"
+    if any(arg_part->arg_part[1].partitioning != arg_part[2], args)
+        return copy_buffered(f, args...)
+    else
+        return f(map(first, args)...)
+    end
+end
+function copy_buffered(f, args...)
+    real_args = map(arg_part->arg_part[1], args)
+    buffered_args = map(arg_part->allocate_copy_buffer(arg_part[2], arg_part[1]), args)
+    for (buf_arg, arg) in zip(buffered_args, real_args)
+        copyto!(buf_arg, arg)
+    end
+    result = f(buffered_args...)
+    for (buf_arg, arg) in zip(buffered_args, real_args)
+        copyto!(arg, buf_arg)
+    end
+    return result
+end
+function allocate_copy_buffer(part::Blocks{N}, A::DArray{T,N}) where {T,N}
+    # FIXME: undef initializer
+    return zeros(part, T, size(A))
+end
+function Base.copyto!(B::DArray{T,N}, A::DArray{T,N}) where {T,N}
+    if size(B) != size(A)
+        throw(DimensionMismatch("Cannot copy from array of size $(size(A)) to array of size $(size(B))"))
+    end
+
+    Bc = B.chunks
+    Ac = A.chunks
+    Asd_all = A.subdomains::DomainBlocks{N}
+
+    Dagger.spawn_datadeps() do
+        for Bidx in CartesianIndices(Bc)
+            Bpart = Bc[Bidx]
+            Bsd = B.subdomains[Bidx]
+
+            # Find the first overlapping subdomain of A
+            if A.partitioning isa Blocks
+                Aidx = CartesianIndex(ntuple(i->fld1(Bsd.indexes[i].start, A.partitioning.blocksize[i]), N))
+            else
+                # Fallback just in case of non-dense partitioning
+                Aidx = first(CartesianIndices(Ac))
+                Asd = first(Asd_all)
+                for dim in 1:N
+                    while Asd.indexes[dim].stop < Bsd.indexes[dim].start
+                        Aidx += CartesianIndex(ntuple(i->i==dim, N))
+                        Asd = Asd_all[Aidx]
+                    end
+                end
+            end
+            Aidx_start = Aidx
+
+            # Find the last overlapping subdomain of A
+            for dim in 1:N
+                while true
+                    Aidx_next = Aidx + CartesianIndex(ntuple(i->i==dim, N))
+                    if !(Aidx_next in CartesianIndices(Ac))
+                        break
+                    end
+                    Asd_next = Asd_all[Aidx_next]
+                    if Asd_next.indexes[dim].start <= Bsd.indexes[dim].stop
+                        Aidx = Aidx_next
+                    else
+                        break
+                    end
+                end
+            end
+            Aidx_end = Aidx
+
+            # Find the span and set of subdomains of A overlapping Bpart
+            Aidx_span = Aidx_start:Aidx_end
+            Asd_view = view(A.subdomains, Aidx_span)
+
+            # Copy all overlapping subdomains of A
+            for Aidx in Aidx_span
+                Asd = Asd_all[Aidx]
+                Apart = Ac[Aidx]
+
+                # Compute the true range
+                range_start = CartesianIndex(ntuple(i->max(Bsd.indexes[i].start, Asd.indexes[i].start), N))
+                range_end = CartesianIndex(ntuple(i->min(Bsd.indexes[i].stop, Asd.indexes[i].stop), N))
+                range_diff = range_end - range_start
+
+                # Compute the offset range into Apart
+                Asd_start = ntuple(i->Asd.indexes[i].start, N)
+                Asd_end = ntuple(i->Asd.indexes[i].stop, N)
+                Arange = range(range_start - CartesianIndex(Asd_start) + CartesianIndex{N}(1),
+                               range_start - CartesianIndex(Asd_start) + CartesianIndex{N}(1) + range_diff)
+
+                # Compute the offset range into Bpart
+                Bsd_start = ntuple(i->Bsd.indexes[i].start, N)
+                Bsd_end = ntuple(i->Bsd.indexes[i].stop, N)
+                Brange = range(range_start - CartesianIndex(Bsd_start) + CartesianIndex{N}(1),
+                               range_start - CartesianIndex(Bsd_start) + CartesianIndex{N}(1) + range_diff)
+
+                # Perform view copy
+                Dagger.@spawn copyto_view!(Out(Bpart), Brange, In(Apart), Arange)
+            end
+        end
+    end
+
+    return B
+end
+function copyto_view!(Bpart, Brange, Apart, Arange)
+    copyto!(view(Bpart, Brange), view(Apart, Arange))
+    return
+end

src/array/darray.jl

@@ -9,15 +9,20 @@ import Serialization: serialize, deserialize
 
 An `N`-dimensional domain over an array.
 """
-struct ArrayDomain{N}
-    indexes::NTuple{N, Any}
+struct ArrayDomain{N,T<:Tuple}
+    indexes::T
 end
-include("../lib/domain-blocks.jl")
-
 
-ArrayDomain(xs...) = ArrayDomain(xs)
+ArrayDomain(xs::T) where T<:Tuple = ArrayDomain{length(xs),T}(xs)
+ArrayDomain(xs::NTuple{N,Base.OneTo}) where N =
+    ArrayDomain{N,NTuple{N,UnitRange{Int}}}(ntuple(i->UnitRange(xs[i]), N))
+ArrayDomain(xs::NTuple{N,Int}) where N =
+    ArrayDomain{N,NTuple{N,UnitRange{Int}}}(ntuple(i->xs[i]:xs[i], N))
+ArrayDomain(xs...) = ArrayDomain((xs...,))
 ArrayDomain(xs::Array) = ArrayDomain((xs...,))
 
+include("../lib/domain-blocks.jl")
+
 indexes(a::ArrayDomain) = a.indexes
 chunks(a::ArrayDomain{N}) where {N} = DomainBlocks(
     ntuple(i->first(indexes(a)[i]), Val(N)), map(x->[length(x)], indexes(a)))
@@ -117,6 +122,7 @@ indicates the number of dimensions in the resulting array.
 """
 Blocks(xs::Int...) = Blocks(xs)
 
+const DArrayDomain{N} = ArrayDomain{N, NTuple{N, UnitRange{Int}}}
 
 """
     DArray{T,N,F}(domain, subdomains, chunks, concat)
@@ -133,8 +139,8 @@ An N-dimensional distributed array of element type T, with a concatenation funct
   and concatenates them along dimension `d`. `cat` is used by default.
 """
 mutable struct DArray{T,N,B<:AbstractBlocks{N},F} <: ArrayOp{T, N}
-    domain::ArrayDomain{N}
-    subdomains::AbstractArray{ArrayDomain{N}, N}
+    domain::DArrayDomain{N}
+    subdomains::AbstractArray{DArrayDomain{N}, N}
     chunks::AbstractArray{Any, N}
     partitioning::B
     concat::F
@@ -143,20 +149,27 @@ mutable struct DArray{T,N,B<:AbstractBlocks{N},F} <: ArrayOp{T, N}
     end
 end
 
+WrappedDArray{T,N} = Union{<:DArray{T,N}, Transpose{<:DArray{T,N}}, Adjoint{<:DArray{T,N}}}
+WrappedDMatrix{T} = WrappedDArray{T,2}
+WrappedDVector{T} = WrappedDArray{T,1}
+DMatrix{T} = DArray{T,2}
+DVector{T} = DArray{T,1}
+
+
 # mainly for backwards-compatibility
 DArray{T, N}(domain, subdomains, chunks, partitioning, concat=cat) where {T,N} =
     DArray(T, domain, subdomains, chunks, partitioning, concat)
 
-function DArray(T, domain::ArrayDomain{N},
-                subdomains::AbstractArray{ArrayDomain{N}, N},
-                chunks::AbstractArray{<:Any, N}, partitioning::B, concat=cat) where {N,B<:AbstractMultiBlocks{N}}
+function DArray(T, domain::DArrayDomain{N},
+                subdomains::AbstractArray{DArrayDomain{N}, N},
+                chunks::AbstractArray{<:Any, N}, partitioning::B, concat=cat) where {N,B<:AbstractBlocks{N}}
     DArray{T,N,B,typeof(concat)}(domain, subdomains, chunks, partitioning, concat)
 end
 
-function DArray(T, domain::ArrayDomain{N},
-                subdomains::ArrayDomain{N},
+function DArray(T, domain::DArrayDomain{N},
+                subdomains::DArrayDomain{N},
                 chunks::Any, partitioning::B, concat=cat) where {N,B<:AbstractSingleBlocks{N}}
-    _subdomains = Array{ArrayDomain{N}, N}(undef, ntuple(i->1, N)...)
+    _subdomains = Array{DArrayDomain{N}, N}(undef, ntuple(i->1, N)...)
     _subdomains[1] = subdomains
     _chunks = Array{Any, N}(undef, ntuple(i->1, N)...)
     _chunks[1] = chunks
@@ -201,6 +214,13 @@ function Base.similar(x::DArray{T,N}) where {T,N}
     return DArray(T, x.domain, x.subdomains, thunks, x.partitioning, x.concat)
 end
 
+function Base.similar(A::DArray{T,N} where T, ::Type{S}, dims::Dims{N}) where {S,N}
+    d = ArrayDomain(map(x->1:x, dims))
+    p = A.partitioning
+    a = AllocateArray(S, (_, _, x...) -> Array{S,N}(undef, x...), d, partition(p, d), p)
+    return _to_darray(a)
+end
+
 Base.copy(x::DArray{T,N,B,F}) where {T,N,B,F} =
     map(identity, x)::DArray{T,N,B,F}
 
@@ -214,7 +234,7 @@ Base.:(/)(x::DArray{T,N,B,F}, y::U) where {T<:Real,U<:Real,N,B,F} =
 A `view` of a `DArray` chunk returns a `DArray` of `Thunk`s.
 """
 function Base.view(c::DArray, d)
-    subchunks, subdomains = lookup_parts(chunks(c), domainchunks(c), d)
+    subchunks, subdomains = lookup_parts(c, chunks(c), domainchunks(c), d)
     d1 = alignfirst(d)
     DArray(eltype(c), d1, subdomains, subchunks, c.partitioning, c.concat)
 end
@@ -252,7 +272,7 @@ function group_indices(cumlength, idxs::AbstractRange)
 end
 
 _cumsum(x::AbstractArray) = length(x) == 0 ? Int[] : cumsum(x)
-function lookup_parts(ps::AbstractArray, subdmns::DomainBlocks{N}, d::ArrayDomain{N}) where N
+function lookup_parts(A::DArray, ps::AbstractArray, subdmns::DomainBlocks{N}, d::ArrayDomain{N}) where N
     groups = map(group_indices, subdmns.cumlength, indexes(d))
     sz = map(length, groups)
     pieces = Array{Any}(undef, sz)
@@ -264,7 +284,15 @@ function lookup_parts(ps::AbstractArray, subdmns::DomainBlocks{N}, d::ArrayDomai
     end
     out_cumlength = map(g->_cumsum(map(x->length(x[2]), g)), groups)
     out_dmn = DomainBlocks(ntuple(x->1,Val(N)), out_cumlength)
-    pieces, out_dmn
+    return pieces, out_dmn
+end
+function lookup_parts(A::DArray, ps::AbstractArray, subdmns::DomainBlocks{N}, d::ArrayDomain{S}) where {N,S}
+    if S != 1
+        throw(BoundsError(A, d.indexes))
+    end
+    inds = CartesianIndices(A)[d.indexes...]
+    new_d = ntuple(i->first(inds).I[i]:last(inds).I[i], N)
+    return lookup_parts(A, ps, subdmns, ArrayDomain(new_d))
 end
 
 """