More tweaks

Project.toml

@@ -53,7 +53,7 @@ Printf = "1"
 Random = "1"
 SafeTestsets = "0.1"
 ScopedValues = "1.3.0"
-Strided = "2"
+Strided = "2.3.3"
 TensorKitSectors = "0.3.5"
 TensorOperations = "5.1"
 Test = "1"

ext/TensorKitCUDAExt/TensorKitCUDAExt.jl

@@ -10,7 +10,7 @@ using TensorKit.Factorizations
 using TensorKit.Strided
 using TensorKit.Factorizations: AbstractAlgorithm
 using TensorKit: SectorDict, tensormaptype, scalar, similarstoragetype, AdjointTensorMap, scalartype, project_symmetric_and_check
-import TensorKit: randisometry, rand, randn
+import TensorKit: randisometry, rand, randn, _copyto!, _add_general_kernel_nonthreaded!, blocktype
 
 using TensorKit: MatrixAlgebraKit
 

ext/TensorKitCUDAExt/cutensormap.jl

@@ -17,6 +17,10 @@ function TensorKit.project_symmetric_and_check(::Type{T}, ::Type{A}, data::Abstr
     return TensorKit.TensorMapWithStorage{T, A}(A(h_t.data), V)
 end
 
+function TensorKit.blocktype(::Type{<:CuTensorMap{T, S}}) where {T, S}
+    return CuMatrix{T, CUDA.DeviceMemory}
+end
+
 for (fname, felt) in ((:zeros, :zero), (:ones, :one))
     @eval begin
         function CUDA.$fname(
@@ -101,18 +105,6 @@ function TensorKit.scalar(t::CuTensorMap{T, S, 0, 0}) where {T, S}
     return isempty(inds) ? zero(scalartype(t)) : @allowscalar @inbounds t.data[only(inds)]
 end
 
-function Base.convert(
-        TT::Type{CuTensorMap{T, S, N₁, N₂}},
-        t::AbstractTensorMap{<:Any, S, N₁, N₂}
-    ) where {T, S, N₁, N₂}
-    if typeof(t) === TT
-        return t
-    else
-        tnew = TT(undef, space(t))
-        return copy!(tnew, t)
-    end
-end
-
 function LinearAlgebra.isposdef(t::CuTensorMap)
     domain(t) == codomain(t) ||
         throw(SpaceMismatch("`isposdef` requires domain and codomain to be the same"))
@@ -138,10 +130,9 @@ function Base.promote_rule(
     return CuTensorMap{T, S, N₁, N₂}
 end
 
-TensorKit.promote_storage_rule(::Type{CuArray{T, N}}, ::Type{<:CuArray{T, N}}) where {T, N} =
+TensorKit.promote_storage_rule(::Type{<:CuArray{T, N}}, ::Type{<:CuArray{T, N}}) where {T, N} =
     CuArray{T, N, CUDA.default_memory}
 
-
 # CuTensorMap exponentation:
 function TensorKit.exp!(t::CuTensorMap)
     domain(t) == codomain(t) ||
@@ -168,3 +159,21 @@ for f in (:sqrt, :log, :asin, :acos, :acosh, :atanh, :acoth)
         return tf
     end
 end
+
+function TensorKit._add_general_kernel_nonthreaded!(
+        tdst::CuTensorMap, tsrc::CuTensorMap, p, transformer::TensorKit.GenericTreeTransformer, α, β, backend...
+    )
+    # preallocate buffers
+    buffers = TensorKit.allocate_buffers(tdst, tsrc, transformer)
+
+    for subtransformer in transformer.data
+        # Special case without intermediate buffers whenever there is only a single block
+        if length(subtransformer[1]) == 1
+            TensorKit._add_transform_single!(tdst, tsrc, p, subtransformer, α, β, backend...)
+        else
+            cu_subtransformer = tuple(CUDA.adapt(CuArray, subtransformer[1]), subtransformer[2:end]...)
+            TensorKit._add_transform_multi!(tdst, tsrc, p, cu_subtransformer, buffers, α, β, backend...)
+        end
+    end
+    return nothing
+end

src/auxiliary/auxiliary.jl

@@ -57,10 +57,12 @@ function _interleave(a::NTuple{N}, b::NTuple{N}) where {N}
     return (a[1], b[1], _interleave(tail(a), tail(b))...)
 end
 
+_copyto!(A, B) = copyto!(A, B)
+
 # Low-overhead implementation of `copyto!` for specific case of `stride(B, 1) < stride(B, 2)`
-# used in indexmanipulations: avoids the overhead of Strided.jl
-function _copyto!(A::StridedView{<:Any, 1}, B::StridedView{<:Any, 2})
-    length(A) == length(B) || throw(DimensionMismatch())
+# for CPU-hosted Arrays # used in indexmanipulations: avoids the overhead of Strided.jl
+function _copyto!(A::StridedView{TA, 1, AA}, B::StridedView{TB, 2, BB}) where {TA <: Number, TB <: Number, AA <: DenseArray{TA}, BB <: DenseArray{TB}}
+    length(A) == length(B) || throw(DimensionMismatch(lazy"length of A ($(length(A))) does not match length of B ($(length(B))"))
 
     Adata = parent(A)
     Astr = stride(A, 1)

src/tensors/abstracttensor.jl

@@ -53,9 +53,11 @@ storagetype(t) = storagetype(typeof(t))
 function storagetype(::Type{T}) where {T <: AbstractTensorMap}
     if T isa Union
         # attempt to be slightly more specific by promoting unions
-        Ma = storagetype(T.a)
-        Mb = storagetype(T.b)
-        return promote_storagetype(Ma, Mb)
+        return promote_storagetype(T.a, T.b)
+    elseif eltype(T) isa Union
+        # attempt to be slightly more specific by promoting unions
+        TU = eltype(T)
+        return promote_storagetype(TU.a, TU.b)
     else
         # fallback definition by using scalartype
         return similarstoragetype(scalartype(T))
@@ -103,8 +105,9 @@ similarstoragetype(X::Type, ::Type{T}) where {T <: Number} =
 
 # implement on tensors
 similarstoragetype(::Type{TT}) where {TT <: AbstractTensorMap} = similarstoragetype(storagetype(TT))
-similarstoragetype(::Type{TT}, ::Type{T}) where {TT <: AbstractTensorMap, T <: Number} =
-    similarstoragetype(storagetype(TT), T)
+function similarstoragetype(::Type{TT}, ::Type{T}) where {TT <: AbstractTensorMap, T <: Number}
+    return similarstoragetype(storagetype(TT), T)
+end
 
 # implement on arrays
 similarstoragetype(::Type{A}) where {A <: DenseVector{<:Number}} = A

src/tensors/braidingtensor.jl

@@ -171,12 +171,15 @@ end
 has_shared_permute(t::BraidingTensor, ::Index2Tuple) = false
 function add_transform!(
         tdst::AbstractTensorMap,
-        tsrc::BraidingTensor, (p₁, p₂)::Index2Tuple,
+        tsrc::BraidingTensor{T, S},
+        (p₁, p₂)::Index2Tuple,
         fusiontreetransform,
         α::Number, β::Number, backend::AbstractBackend...
-    )
+    ) where {T, S}
+    tsrc_map = similar(tdst, storagetype(tdst), space(tsrc))
+    copy!(tsrc_map, tsrc)
     return add_transform!(
-        tdst, TensorMap(tsrc), (p₁, p₂), fusiontreetransform, α, β,
+        tdst, tsrc_map, (p₁, p₂), fusiontreetransform, α, β,
         backend...
     )
 end

src/tensors/treetransformers.jl

@@ -46,7 +46,7 @@ function AbelianTreeTransformer(transform, p, Vdst, Vsrc)
 end
 
 const _GenericTransformerData{T, N} = Tuple{
-    Matrix{T},
+    DenseMatrix{T},
     Tuple{NTuple{N, Int}, Vector{Tuple{NTuple{N, Int}, Int}}},
     Tuple{NTuple{N, Int}, Vector{Tuple{NTuple{N, Int}, Int}}},
 }

test/cuda/tensors.jl

@@ -98,8 +98,8 @@ for V in spacelist
                 next = @constinferred Nothing iterate(bs, state)
                 b2 = @constinferred block(t, first(blocksectors(t)))
                 @test b1 == b2
-                @test_broken eltype(bs) === Pair{typeof(c), typeof(b1)}
-                @test_broken typeof(b1) === TensorKit.blocktype(t)
+                @test eltype(bs) === Pair{typeof(c), typeof(b1)}
+                @test typeof(b1) === TensorKit.blocktype(t)
                 @test typeof(c) === sectortype(t)
             end
         end
@@ -162,8 +162,8 @@ for V in spacelist
                 next = @constinferred Nothing iterate(bs, state)
                 b2 = @constinferred block(t', first(blocksectors(t')))
                 @test b1 == b2
-                @test_broken eltype(bs) === Pair{typeof(c), typeof(b1)}
-                @test_broken typeof(b1) === TensorKit.blocktype(t')
+                @test eltype(bs) === Pair{typeof(c), typeof(b1)}
+                @test typeof(b1) === TensorKit.blocktype(t')
                 @test typeof(c) === sectortype(t)
                 # linear algebra
                 @test isa(@constinferred(norm(t)), real(T))