Add a disamgiguating conversion

src/tensors/abstractblocktensor/conversion.jl

@@ -1,37 +1,50 @@
 # Conversion
 # ----------
-function Base.convert(::Type{TensorMap}, t::AbstractBlockTensorMap)
-    S = spacetype(t)
-    N₁, N₂ = numout(t), numin(t)
-    cod = ProductSpace{S, N₁}(oplus.(codomain(t).spaces))
-    dom = ProductSpace{S, N₂}(oplus.(domain(t).spaces))
-    tdst = similar(t, cod ← dom)
-
-    issparse(t) && zerovector!(tdst)
 
+function _copy_subblocks!(tdst, tsrc)
+    S = spacetype(tsrc)
+    N₁, N₂ = numout(tsrc), numin(tsrc)
     for ((f₁, f₂), arr) in subblocks(tdst)
         blockax = ntuple(N₁ + N₂) do i
             return if i <= N₁
-                blockedrange(map(Base.Fix2(dim, f₁.uncoupled[i]), space(t, i)))
+                blockedrange(map(Base.Fix2(dim, f₁.uncoupled[i]), space(tsrc, i)))
             else
-                blockedrange(map(Base.Fix2(dim, f₂.uncoupled[i - N₁]), space(t, i)'))
+                blockedrange(map(Base.Fix2(dim, f₂.uncoupled[i - N₁]), space(tsrc, i)'))
             end
         end
 
-        for (k, v) in nonzero_pairs(t)
+        for (k, v) in nonzero_pairs(tsrc)
             indices = getindex.(blockax, Block.(Tuple(k)))
             arr_slice = arr[indices...]
             # need to check for empty since fusion tree pair might not be present
             isempty(arr_slice) || copy!(arr_slice, v[f₁, f₂])
         end
     end
+    return tdst
+end
 
+function Base.convert(::Type{TensorMap}, t::AbstractBlockTensorMap)
+    S = spacetype(t)
+    N₁, N₂ = numout(t), numin(t)
+    cod = ProductSpace{S, N₁}(oplus.(codomain(t).spaces))
+    dom = ProductSpace{S, N₂}(oplus.(domain(t).spaces))
+    tdst = TensorMapWithStorageType{scalartype(t), storagetype(t)}(undef, cod, dom)
+
+    issparse(t) && zerovector!(tdst)
+    _copy_subblocks!(tdst, t)
     return tdst
 end
 
-function Base.convert(::Type{T}, t::AbstractBlockTensorMap) where {T <: TensorMap}
-    tdst = convert(TensorMap, t)
-    return convert(T, tdst)
+function Base.convert(::Type{TT}, t::AbstractBlockTensorMap) where {TT <: TensorMap}
+    S = spacetype(t)
+    N₁, N₂ = numout(t), numin(t)
+    cod = ProductSpace{S, N₁}(oplus.(codomain(t).spaces))
+    dom = ProductSpace{S, N₂}(oplus.(domain(t).spaces))
+    tdst = TT(undef, cod ← dom)
+    issparse(t) && zerovector!(tdst)
+
+    _copy_subblocks!(tdst, t)
+    return tdst
 end
 
 function Base.convert(::Type{TT}, t::AbstractTensorMap) where {TT <: AbstractBlockTensorMap}

src/tensors/tensoroperations.jl

@@ -15,6 +15,17 @@ function TO.tensoradd_type(TC, A::AdjointBlockTensorMap, pA::Index2Tuple, conjA:
     return TO.tensoradd_type(TC, A', adjointtensorindices(A, pA), !conjA)
 end
 
+# copy blocks back to CPU/collect them into an array
+# seems necessary for GPU-backed BlockTensorMaps but
+# maybe not the most efficient approach?
+function TO.tensorscalar(t::AbstractBlockTensorMap{T, S, 0, 0}) where {T, S}
+    Bs = TK.blocks(t)
+    B_ends = collect.(map(b -> collect.(getfield(b, :blocks)), map(last, Bs)))
+    inds = findall(!iszero ∘ last, B_ends)
+    isempty(inds) && return zero(TKscalartype(t))
+    return only(last(B_ends[only(inds)]))
+end
+
 # tensoralloc_contract
 # --------------------
 for TTA in (:AbstractTensorMap, :AbstractBlockTensorMap), TTB in (:AbstractTensorMap, :AbstractBlockTensorMap)