update cuda patch

Author: GiggleLiu

src/cuda.jl

@@ -1,9 +1,21 @@
 using .CUDA
+using TropicalGEMM: XTranspose, NativeTypes, Tropical
+using LinearAlgebra
 
 function onehotmask(A::CuArray{T}, X::CuArray{T}) where T
     mask = X .== inv.(A)
     ci = argmax(mask)
     mask .= false
     mask[CuArray([ci])] = true
     return mask
+end
+
+# fix the matrix multiplication ambiguity
+const CTranspose{T} = Transpose{T, <:StridedCuVecOrMat}
+for (TA, CTA) in [(:AbstractMatrix, :CuMatrix), (:XTranspose, :CTranspose)]
+    for (TB, CTB) in [(:AbstractMatrix, :CuMatrix), (:XTranspose, :CTranspose)]
+        @eval function LinearAlgebra.mul!(o::CuMatrix{T}, a::$TA{T}, b::$TB{T}, α::Number, β::Number) where {T<:Tropical{<:NativeTypes}}
+            invoke(LinearAlgebra.mul!, Tuple{CuMatrix, $CTA, $CTB, Number, Number}, o, a, b, α, β)
+        end
+    end
 end

src/networks.jl

@@ -109,6 +109,7 @@ Optimize the contraction order.
     Check `optimize_kahypar` method in package `OMEinsumContractionOrders`.
     * `:auto`, also the kahypar + greedy approach, but determines `sc_target` automatically. It is slower!
     * `:greedy`, the greedy approach. Check `optimize_greedy` in package `OMEinsum`.
+    * `:tree`, the approach of running simulated annealing on expression tree. Check `optimize_tree` in package `OMEinsumContractionOrders`.
     * `:sa`, the simulated annealing approach. Check `optimize_sa` in package `OMEinsumContractionOrders`.
     * `:raw`, do nothing and return the raw EinCode.
 """
@@ -162,4 +163,4 @@ function set_packing(sets; kwargs...)
     n = length(sets)
     code = EinCode(([(i,) for i=1:n]..., [(i,j) for i=1:n,j=1:n if j>i && !isempty(sets[i] ∩ sets[j])]...), ())
     Independence(optimize_code(code; kwargs...))
-end
+end