|
| 1 | +export ParallelBlocks |
| 2 | + |
| 3 | +using Statistics |
| 4 | + |
| 5 | +struct ParallelBlocks{N} <: Dagger.AbstractSingleBlocks{N} |
| 6 | + n::Int |
| 7 | +end |
| 8 | +ParallelBlocks(n::Integer) = ParallelBlocks{0}(n) |
| 9 | +ParallelBlocks{N}(dist::ParallelBlocks) where N = ParallelBlocks{N}(dist.n) |
| 10 | +ParallelBlocks() = ParallelBlocks(Dagger.num_processors()) |
| 11 | + |
| 12 | +Base.convert(::Type{ParallelBlocks{N}}, dist::ParallelBlocks) where N = |
| 13 | + ParallelBlocks{N}(dist.n) |
| 14 | + |
| 15 | +wrap_chunks(chunks::Vector{<:Dagger.Chunk}, N::Integer, dist::ParallelBlocks) = |
| 16 | + wrap_chunks(chunks, N, dist.n) |
| 17 | +wrap_chunks(chunks::Vector{<:Dagger.Chunk}, N::Integer, n::Integer) = |
| 18 | + convert(Array{Any}, reshape(chunks, ntuple(i->i == 1 ? n : 1, N))) |
| 19 | + |
| 20 | +function _finish_allocation(f::Function, dist::ParallelBlocks, dims::NTuple{N,Int}) where N |
| 21 | + d = ArrayDomain(map(x->1:x, dims)) |
| 22 | + s = reshape([d for _ in 1:dist.n], |
| 23 | + ntuple(i->i == 1 ? dist.n : 1, N)) |
| 24 | + data = [f(dims) for _ in 1:dist.n] |
| 25 | + dist = ParallelBlocks{N}(dist) |
| 26 | + chunks = wrap_chunks(map(Dagger.tochunk, data), N, dist) |
| 27 | + return Dagger.DArray(eltype(first(data)), d, s, chunks, dist) |
| 28 | +end |
| 29 | + |
| 30 | +for fn in [:rand, :randn, :zeros, :ones] |
| 31 | + @eval begin |
| 32 | + function Base.$fn(dist::ParallelBlocks, ::Type{ET}, dims::Dims) where {ET} |
| 33 | + f(block) = $fn(ET, block) |
| 34 | + _finish_allocation(f, dist, dims) |
| 35 | + end |
| 36 | + Base.$fn(dist::ParallelBlocks, T::Type, dims::Integer...) = $fn(dist, T, dims) |
| 37 | + Base.$fn(dist::ParallelBlocks, T::Type, dims::Tuple) = $fn(dist, T, dims) |
| 38 | + Base.$fn(dist::ParallelBlocks, dims::Integer...) = $fn(dist, Float64, dims) |
| 39 | + Base.$fn(dist::ParallelBlocks, dims::Tuple) = $fn(dist, Float64, dims) |
| 40 | + end |
| 41 | +end |
| 42 | +# FIXME: sprand |
| 43 | + |
| 44 | +function Dagger.distribute(data::AbstractArray{T,N}, dist::ParallelBlocks) where {T,N} |
| 45 | + dims = size(data) |
| 46 | + d = ArrayDomain(map(x->1:x, dims)) |
| 47 | + s = Dagger.DomainBlocks(ntuple(_->1, N), |
| 48 | + ntuple(i->[dims[i]], N)) |
| 49 | + chunks = [Dagger.tochunk(copy(data)) for _ in 1:dist.n] |
| 50 | + new_dist = ParallelBlocks{N}(dist) |
| 51 | + return Dagger.DArray(T, d, s, wrap_chunks(chunks, N, dist), new_dist) |
| 52 | +end |
| 53 | + |
| 54 | +_invalid_call_pblocks(f::Symbol) = |
| 55 | + error("`$f` is not valid for a `DArray` partitioned with `ParallelBlocks`.\nConsider `Dagger.pmap($f, x)` instead.") |
| 56 | + |
| 57 | +Base.collect(::Dagger.DArray{T,N,<:ParallelBlocks} where {T,N}) = |
| 58 | + _invalid_call_pblocks(:collect) |
| 59 | +Base.getindex(::Dagger.DArray{T,N,<:ParallelBlocks} where {T,N}, x...) = |
| 60 | + _invalid_call_pblocks(:getindex) |
| 61 | +Base.setindex!(::Dagger.DArray{T,N,<:ParallelBlocks} where {T,N}, value, x...) = |
| 62 | + _invalid_call_pblocks(:setindex!) |
| 63 | + |
| 64 | +function pmap(f::Function, A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} |
| 65 | + # TODO: Chunks might not be `Array`s |
| 66 | + # FIXME |
| 67 | + #AT = Array{T,N} |
| 68 | + #ET = eltype(Base.promote_op(f, AT)) |
| 69 | + ET = Any |
| 70 | + new_chunks = map(A.chunks) do chunk |
| 71 | + Dagger.@spawn f(chunk) |
| 72 | + end |
| 73 | + return DArray(ET, A.domain, A.subdomains, new_chunks, A.partitioning) |
| 74 | +end |
| 75 | +# FIXME: More useful `show` method |
| 76 | +Base.show(io::IO, ::MIME"text/plain", A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} = |
| 77 | + print(io, typeof(A)) |
| 78 | +pfetch(A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} = |
| 79 | + map(fetch, A.chunks) |
| 80 | +pcollect(A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} = |
| 81 | + map(collect, pfetch(A)) |
| 82 | + |
| 83 | +function Base.map(f::Function, A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} |
| 84 | + ET = Base.promote_op(f, T) |
| 85 | + new_chunks = map(A.chunks) do chunk |
| 86 | + Dagger.@spawn map(f, chunk) |
| 87 | + end |
| 88 | + return DArray(ET, A.domain, A.subdomains, new_chunks, A.partitioning) |
| 89 | +end |
| 90 | +function Base.map!(f::Function, |
| 91 | + x::Dagger.DArray{T1,N1,ParallelBlocks{N1}} where {T1,N1}, |
| 92 | + y::Dagger.DArray{T2,N2,ParallelBlocks{N2}} where {T2,N2}) |
| 93 | + x_dist = x.partitioning |
| 94 | + y_dist = y.partitioning |
| 95 | + if x_dist.n != y_dist.n |
| 96 | + throw(ArgumentError("Can't `map!` over non-matching `ParallelBlocks` distributions: $(x_dist.n) != $(y_dist.n)")) |
| 97 | + end |
| 98 | + @sync for i in 1:x_dist.n |
| 99 | + Dagger.@spawn map!(f, x.chunks[i], y.chunks[i]) |
| 100 | + end |
| 101 | +end |
| 102 | + |
| 103 | +#= |
| 104 | +function Base.reduce(f::Function, x::Dagger.DArray{T,N,ParallelBlocks{N}}; |
| 105 | + dims=:) where {T,N} |
| 106 | + error("Out-of-place Reduce") |
| 107 | + if dims == Base.:(:) |
| 108 | + localpart = fetch(Dagger.reduce_async(f, x)) |
| 109 | + return MPI.Allreduce(localpart, f, comm) |
| 110 | + elseif dims === nothing |
| 111 | + localpart = fetch(x.chunks[1]) |
| 112 | + return MPI.Allreduce(localpart, f, comm) |
| 113 | + else |
| 114 | + error("Not yet implemented") |
| 115 | + end |
| 116 | +end |
| 117 | +=# |
| 118 | +function allreduce!(op::Function, x::Dagger.DArray{T,N,ParallelBlocks{N}}; nchunks::Integer=0) where {T,N} |
| 119 | + if nchunks == 0 |
| 120 | + nchunks = x.partitioning.n |
| 121 | + end |
| 122 | + @assert nchunks == x.partitioning.n "Number of chunks must match the number of partitions" |
| 123 | + |
| 124 | + # Split each chunk along the last dimension |
| 125 | + chunk_size = cld(size(x, ndims(x)), nchunks) |
| 126 | + chunk_dist = Blocks(ntuple(i->i == N ? chunk_size : size(x, i), N)) |
| 127 | + chunk_ds = partition(chunk_dist, x.subdomains[1]) |
| 128 | + num_par_chunks = length(x.chunks) |
| 129 | + |
| 130 | + # Allocate temporary buffer |
| 131 | + y = copy(x) |
| 132 | + |
| 133 | + # Ring-reduce into temporary buffer |
| 134 | + Dagger.spawn_datadeps() do |
| 135 | + for j in 1:length(chunk_ds) |
| 136 | + for i in 1:num_par_chunks |
| 137 | + for step in 1:(num_par_chunks-1) |
| 138 | + from_idx = i |
| 139 | + to_idx = mod1(i+step, num_par_chunks) |
| 140 | + from_chunk = x.chunks[from_idx] |
| 141 | + to_chunk = y.chunks[to_idx] |
| 142 | + sd = chunk_ds[mod1(j+i-1, length(chunk_ds))].indexes |
| 143 | + # FIXME: Specify aliasing based on `sd` |
| 144 | + Dagger.@spawn _reduce_view!(op, |
| 145 | + InOut(to_chunk), sd, |
| 146 | + In(from_chunk), sd) |
| 147 | + end |
| 148 | + end |
| 149 | + end |
| 150 | + |
| 151 | + # Copy from temporary buffer back to origin |
| 152 | + for i in 1:num_par_chunks |
| 153 | + Dagger.@spawn copyto!(Out(x.chunks[i]), In(y.chunks[i])) |
| 154 | + end |
| 155 | + end |
| 156 | + |
| 157 | + return x |
| 158 | +end |
| 159 | +function _reduce_view!(op, to, to_view, from, from_view) |
| 160 | + to_viewed = view(to, to_view...) |
| 161 | + from_viewed = view(from, from_view...) |
| 162 | + reduce!(op, to_viewed, from_viewed) |
| 163 | + return |
| 164 | +end |
| 165 | +function reduce!(op, to, from) |
| 166 | + to .= op.(to, from) |
| 167 | +end |
| 168 | + |
| 169 | +function Statistics.mean!(A::Dagger.DArray{T,N,ParallelBlocks{N}}) where {T,N} |
| 170 | + allreduce!(+, A) |
| 171 | + len = length(A.chunks) |
| 172 | + map!(x->x ./ len, A, A) |
| 173 | + return A |
| 174 | +end |
0 commit comments