Add test for permutedims bug

NDTensors/src/lib/BlockSparseArrays/src/abstractblocksparsearray/wrappedabstractblocksparsearray.jl

@@ -273,3 +273,22 @@ function Base.similar(
 )
   return blocksparse_similar(a, elt, axes)
 end
+
+# TODO: Implement this in a more generic way using a smarter `copyto!`,
+# which is ultimately what `Array{T,N}(::AbstractArray{<:Any,N})` calls.
+# These are defined for now to avoid scalar indexing issues when there
+# are blocks on GPU.
+function Base.Array{T,N}(a::BlockSparseArrayLike{<:Any,N}) where {T,N}
+  # First make it dense, then move to CPU.
+  # Directly copying to CPU causes some issues with
+  # scalar indexing on GPU which we have to investigate.
+  a_dest = similartype(blocktype(a), T)(undef, size(a))
+  a_dest .= a
+  return Array{T,N}(a_dest)
+end
+function Base.Array{T}(a::BlockSparseArrayLike) where {T}
+  return Array{T,ndims(a)}(a)
+end
+function Base.Array(a::BlockSparseArrayLike)
+  return Array{eltype(a)}(a)
+end

NDTensors/src/lib/BlockSparseArrays/test/test_basics.jl

@@ -297,6 +297,15 @@ using .NDTensorsTestUtils: devices_list, is_supported_eltype
     @test block_nstored(b) == 2
     @test nstored(b) == 2 * 4 + 3 * 3
 
+    a = dev(BlockSparseArray{elt}([1, 1, 1], [1, 2, 3], [2, 2, 1], [1, 2, 1]))
+    a[Block(3, 2, 2, 3)] = dev(randn(1, 2, 2, 1))
+    perm = (2, 3, 4, 1)
+    for b in (PermutedDimsArray(a, perm), permutedims(a, perm))
+      @test Array(b) == permutedims(Array(a), perm)
+      @test issetequal(block_stored_indices(b), [Block(2, 2, 3, 3)])
+      @test @allowscalar b[Block(2, 2, 3, 3)] == permutedims(a[Block(3, 2, 2, 3)], perm)
+    end
+
     a = BlockSparseArray{elt}(undef, ([2, 3], [3, 4]))
     @views for b in [Block(1, 2), Block(2, 1)]
       a[b] = randn(elt, size(a[b]))