fmt option in adjaceny matrix + propagate copy_xj for Metal

GNNGraphs/src/convert.jl

@@ -207,7 +207,7 @@ function to_sparse(A::ADJMAT_T, T = nothing; dir = :out, num_nodes = nothing,
         A = sparse(A)
     end
     if !weighted
-        A = map(x -> ifelse(x > 0, T(1), T(0)), A)
+        A = binarize(A, T)
     end
     return A, num_nodes, num_edges
 end

GNNGraphs/src/query.jl

@@ -218,7 +218,7 @@ end
 adjacency_list(g::GNNGraph; dir = :out) = adjacency_list(g, 1:(g.num_nodes); dir)
 
 """
-    adjacency_matrix(g::GNNGraph, T=eltype(g); dir=:out, weighted=true)
+    adjacency_matrix(g::GNNGraph, T=eltype(g); dir=:out, weighted=true, fmt=nothing)
 
 Return the adjacency matrix `A` for the graph `g`. 
 
@@ -227,29 +227,39 @@ If `dir=:in` instead, `A[i,j] > 0` denotes the presence of an edge from node `j`
 
 User may specify the eltype `T` of the returned matrix. 
 
-If `weighted=true`, the `A` will contain the edge weights if any, otherwise the elements of `A` will be either 0 or 1.
-"""
-function Graphs.adjacency_matrix(g::GNNGraph{<:COO_T}, T::DataType = eltype(g); dir = :out,
-                                 weighted = true)
-    A, n, m = to_sparse(g.graph, T; num_nodes = g.num_nodes, weighted)
+If `weighted=true`, the matrix `A` will contain the edge weights if any, otherwise the elements of `A`.
+If the graph does not contain edge weights, or if `weighted=false`, the adjacency matrix will contain only 0s and 1s.
+
+The argument `fmt` can be used to specify the desired format of the returned matrix. Possible values are:
+- `nothing`: return the matrix in the same format as the underlying graph representation.
+- `:sparse`: return a sparse matrix (default for COO graphs).
+- `:dense`: return a dense matrix (default for adjacency matrix graphs).
+"""
+function Graphs.adjacency_matrix(g::GNNGraph, T::DataType = eltype(g); dir = :out,
+                                 weighted = true, fmt = nothing)
+    if fmt === nothing
+        if g.graph isa COO_T
+            fmt = :sparse
+        elseif g.graph isa SPARSE_T
+            fmt = :sparse
+        else
+            fmt = :dense
+        end
+    end
+    @assert fmt ∈ [:sparse, :dense]
+    if fmt == :sparse
+        A, n, m = to_sparse(g.graph, T; num_nodes = g.num_nodes, weighted)
+    else
+        A, n, m = to_dense(g.graph, T; num_nodes = g.num_nodes, weighted)
+    end
     @assert size(A) == (n, n)
     return dir == :out ? A : A'
 end
 
-function Graphs.adjacency_matrix(g::GNNGraph{<:ADJMAT_T}, T::DataType = eltype(g);
-                                 dir = :out, weighted = true)
-    @assert dir ∈ [:in, :out]
-    A = g.graph
-    if !weighted
-        A = binarize(A, T)
-    end
-    A = T != eltype(A) ? T.(A) : A
-    return dir == :out ? A : A'
-end
 
 function CRC.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
-            dir = :out, weighted = true) where {G <: GNNGraph{<:ADJMAT_T}}
-    A = adjacency_matrix(g, T; dir, weighted)
+            dir=:out, weighted=true, fmt=nothing) where {G <: GNNGraph{<:ADJMAT_T}}
+    A = adjacency_matrix(g, T; dir, weighted, fmt)
     if !weighted
         function adjacency_matrix_pullback_noweight(Δ)
             return (CRC.NoTangent(), CRC.ZeroTangent(), CRC.NoTangent())  
@@ -266,8 +276,8 @@ function CRC.rrule(::typeof(adjacency_matrix), g::G, T::DataType;
 end
 
 function CRC.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
-            dir = :out, weighted = true) where {G <: GNNGraph{<:COO_T}}
-    A = adjacency_matrix(g, T; dir, weighted)
+            dir=:out, weighted=true, fmt=nothing) where {G <: GNNGraph{<:COO_T}}
+    A = adjacency_matrix(g, T; dir, weighted, fmt)
     w = get_edge_weight(g)
     if !weighted || w === nothing
         function adjacency_matrix_pullback_noweight(Δ)

GNNGraphs/test/Project.toml

@@ -2,7 +2,6 @@
 FiniteDifferences = "26cc04aa-876d-5657-8c51-4c34ba976000"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GNNGraphs = "aed8fd31-079b-4b5a-b342-a13352159b8c"
-GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
@@ -20,6 +19,3 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 TestItems = "1c621080-faea-4a02-84b6-bbd5e436b8fe"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
-
-[compat]
-GPUArraysCore = "0.1"

GNNlib/Project.toml

@@ -16,10 +16,12 @@ Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 [weakdeps]
 AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+Metal = "dde4c033-4e86-420c-a63e-0dd931031962"
 
 [extensions]
 GNNlibAMDGPUExt = "AMDGPU"
 GNNlibCUDAExt = "CUDA"
+GNNlibMetalExt = "Metal"
 
 [compat]
 AMDGPU = "1"
@@ -28,6 +30,7 @@ ChainRulesCore = "1.24"
 DataStructures = "0.18"
 GNNGraphs = "1.4"
 LinearAlgebra = "1"
+Metal = "1.0"
 MLUtils = "0.4"
 NNlib = "0.9"
 Random = "1"

GNNlib/ext/GNNlibCUDAExt.jl

@@ -15,7 +15,14 @@ const CUDA_COO_T = Tuple{T, T, V} where {T <: AnyCuArray{<:Integer}, V <: Union{
 ## avoid the fast path on gpu until we have better cuda support
 function GNNlib.propagate(::typeof(copy_xj), g::GNNGraph{<:COO_T}, ::typeof(+),
         xi, xj::AnyCuMatrix, e)
-    A = _adjacency_matrix(g, eltype(xj); weighted = false)
+
+    if !g.is_coalesced
+        # Revisit after 
+        # https://github.com/JuliaGPU/CUDA.jl/issues/1113
+        A = adjacency_matrix(g, eltype(xj); weighted=false, fmt=:dense)
+    else
+        A = adjacency_matrix(g, eltype(xj); weighted=false, fmt=:sparse)
+    end
 
     return xj * A
 end
@@ -47,21 +54,4 @@ end
 
 # Flux.Zygote.@nograd compute_degree
 
-## CUSTOM ADJACENCY_MATRIX IMPLEMENTATION FOR CUDA COO GRAPHS, returning dense matrix when not coalesced, more efficient 
-
-function _adjacency_matrix(g::GNNGraph{<:CUDA_COO_T}, T::DataType = eltype(g); dir = :out,
-                                 weighted = true)
-    if !g.is_coalesced
-        # Revisit after 
-        # https://github.com/JuliaGPU/CUDA.jl/issues/1113
-        A, n, m = to_dense(g.graph, T; num_nodes = g.num_nodes, weighted) # if not coalesced, construction of sparse matrix is slow
-    else
-        A, n, m = to_sparse(g.graph, T; num_nodes = g.num_nodes, weighted, is_coalesced = true)
-    end
-    @assert size(A) == (n, n)
-    return dir == :out ? A : A'
-end
-
-@non_differentiable _adjacency_matrix(x...)
-
 end #module

GNNlib/ext/GNNlibMetalExt.jl

@@ -0,0 +1,23 @@
+module GNNlibMetalExt
+
+using Metal
+using Random, Statistics, LinearAlgebra
+using GNNlib: GNNlib, propagate, copy_xj, e_mul_xj, w_mul_xj
+using GNNGraphs: GNNGraph, COO_T, SPARSE_T, adjacency_matrix
+using ChainRulesCore: @non_differentiable
+
+const METAL_COO_T = Tuple{T, T, V} where {T <: MtlVector{<:Integer}, V <: Union{Nothing, MtlVector}}
+
+###### PROPAGATE SPECIALIZATIONS ####################
+
+## COPY_XJ 
+
+## Metal does not support sparse arrays yet and neither scater.
+## Have to use dense adjacency matrix multiplication for now.
+function GNNlib.propagate(::typeof(copy_xj), g::GNNGraph{<:METAL_COO_T}, ::typeof(+),
+        xi, xj::AbstractMatrix, e)
+    A = adjacency_matrix(g, eltype(xj), weighted=false, fmt=:dense)
+    return xj * A
+end
+
+end #module

GNNlib/test/Project.toml

@@ -4,7 +4,6 @@ Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GNNGraphs = "aed8fd31-079b-4b5a-b342-a13352159b8c"
 GNNlib = "a6a84749-d869-43f8-aacc-be26a1996e48"
-GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
@@ -17,6 +16,3 @@ Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
-
-[compat]
-GPUArraysCore = "0.1"

GraphNeuralNetworks/docs/src/index.md

@@ -30,7 +30,7 @@ pkg> add GraphNeuralNetworks
 
 Let's give a brief overview of the package by solving a graph regression problem with synthetic data. 
 
-Other usage examples can be found in the [examples](https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/tree/master/GraphNeuralNetworks/examples) folder, in the [notebooks](https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/tree/master/GraphNeuralNetworks/notebooks) folder, and in the [tutorials](https://juliagraphs.org/GraphNeuralNetworks.jl/tutorials/) section of the documentation.
+Other usage examples can be found in the [examples](https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/tree/master/GraphNeuralNetworks/examples) folder, in the [notebooks](https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/tree/master/GraphNeuralNetworks/notebooks) folder, and in the [tutorials](https://juliagraphs.org/GraphNeuralNetworks.jl/docs/GraphNeuralNetworks.jl/dev/tutorials/gnn_intro/) section of the documentation.
 
 ### Data preparation
 

GraphNeuralNetworks/test/Project.toml

@@ -5,10 +5,10 @@ Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GNNGraphs = "aed8fd31-079b-4b5a-b342-a13352159b8c"
 GNNlib = "a6a84749-d869-43f8-aacc-be26a1996e48"
-GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 GraphNeuralNetworks = "cffab07f-9bc2-4db1-8861-388f63bf7694"
 Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
+Metal = "dde4c033-4e86-420c-a63e-0dd931031962"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
@@ -17,5 +17,3 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
-[compat]
-GPUArraysCore = "0.1"