More tweaks

Author: kshyatt

Project.toml

@@ -55,7 +55,7 @@ Printf = "1"
 Random = "1"
 SafeTestsets = "0.1"
 ScopedValues = "1.3.0"
-Strided = "2"
+Strided = "2.3.4"
 TensorKitSectors = "0.3.6"
 TensorOperations = "5.1"
 Test = "1"
@@ -89,3 +89,6 @@ cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 
 [targets]
 test = ["ArgParse", "Adapt", "Aqua", "AllocCheck", "Combinatorics", "CUDA", "cuTENSOR", "GPUArrays", "JET", "LinearAlgebra", "SafeTestsets", "TensorOperations", "Test", "TestExtras", "ChainRulesCore", "ChainRulesTestUtils", "FiniteDifferences", "Zygote", "Mooncake"]
+
+[sources]
+Strided = {url = "https://github.com/QuantumKitHub/Strided.jl", rev = "ksh/copyto"}

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

@@ -6,6 +6,9 @@ const AdjointCuTensorMap{T, S, N₁, N₂} = AdjointTensorMap{T, S, N₁, N₂,
 function CuTensorMap(t::TensorMap{T, S, N₁, N₂, A}) where {T, S, N₁, N₂, A}
     return CuTensorMap{T, S, N₁, N₂}(CuArray{T}(t.data), space(t))
 end
+function TensorMap{T, S, N₁, N₂, DA}(t::TensorMap{T, S, N₁, N₂, HA}) where {T, S, N₁, N₂, DA <: CuArray{T}, HA <: Array{T}}
+    return CuTensorMap{T, S, N₁, N₂}(CuArray{T}(t.data), space(t))
+end
 
 # project_symmetric! doesn't yet work for GPU types, so do this on the host, then copy
 function TensorKit.project_symmetric_and_check(::Type{T}, ::Type{A}, data::AbstractArray, V::TensorMapSpace; tol = sqrt(eps(real(float(eltype(data)))))) where {T, A <: CuVector{T}}
@@ -101,18 +104,6 @@ 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 +129,9 @@ 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 +158,21 @@ for f in (:sqrt, :log, :asin, :acos, :acosh, :atanh, :acoth)
         return tf
     end
 end
+
+function TensorKit.add_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/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))
@@ -103,11 +105,19 @@ 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
+function similarstoragetype(::Type{<:AbstractTensorMap{T, S, N₁, N₂}}, ::Type{TA}) where {T <: Number, TA <: DenseVector, S, N₁, N₂}
+    return similarstoragetype(TA, T)
+end
+function similarstoragetype(t::AbstractTensorMap{T, S, N₁, N₂}, ::Type{TA}) where {T <: Number, TA <: DenseVector, S, N₁, N₂}
+    return similarstoragetype(typeof(t), TA)
+end
 
 # implement on arrays
 similarstoragetype(::Type{A}) where {A <: DenseVector{<:Number}} = A
+similarstoragetype(::Type{A}, ::Type{A}) where {A <: DenseVector{<:Number}} = A
 Base.@assume_effects :foldable similarstoragetype(::Type{A}) where {A <: AbstractArray{<:Number}} =
     Core.Compiler.return_type(similar, Tuple{A, Int})
 Base.@assume_effects :foldable similarstoragetype(::Type{A}, ::Type{T}) where {A <: AbstractArray, T <: Number} =

src/tensors/adjoint.jl

@@ -22,6 +22,8 @@ Base.adjoint(t::AbstractTensorMap) = AdjointTensorMap(t)
 space(t::AdjointTensorMap) = adjoint(space(parent(t)))
 dim(t::AdjointTensorMap) = dim(parent(t))
 storagetype(::Type{AdjointTensorMap{T, S, N₁, N₂, TT}}) where {T, S, N₁, N₂, TT} = storagetype(TT)
+similarstoragetype(::AdjointTensorMap{T, S, N₁, N₂, TT}, ::Type{T′}) where {T, S, N₁, N₂, TT, T′ <: Number} = similarstoragetype(TT, T′)
+similarstoragetype(::AdjointTensorMap{T, S, N₁, N₂, TT}, ::Type{TA}) where {T, S, N₁, N₂, TT, TA <: DenseVector} = similarstoragetype(TT, TA)
 
 # Blocks and subblocks
 #----------------------

src/tensors/braidingtensor.jl

@@ -145,12 +145,10 @@ function block(b::BraidingTensor, s::Sector)
     # TODO: probably always square?
     m = blockdim(codomain(b), s)
     n = blockdim(domain(b), s)
-    data = Matrix{eltype(b)}(undef, (m, n))
+    data = zeros(eltype(b), (m, n))
 
     length(data) == 0 && return data # s ∉ blocksectors(b)
 
-    data = fill!(data, zero(eltype(b)))
-
     V1, V2 = codomain(b)
     if sectortype(b) === Trivial
         d1, d2 = dim(V1), dim(V2)
@@ -182,12 +180,15 @@ 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
@@ -287,11 +288,15 @@ function planarcontract!(
         backend, allocator
     )
     # special case only defined for contracting 2 indices
-    length(oindB) == length(cindB) == 2 ||
+    if !(length(oindB) == length(cindB) == 2)
+        # horrible!!!!!
+        tB′ = TensorMap(B)
+        tB = TensorMapWithStorage{eltype(B), similarstoragetype(A, eltype(B)), spacetype(tB′), numout(tB′), numin(tB′)}(tB′)
         return planarcontract!(
-        C, A, (oindA, cindA), TensorMap(B), (cindB, oindB), (p1, p2),
-        α, β, backend, allocator
-    )
+            C, A, (oindA, cindA), tB, (cindB, oindB), (p1, p2),
+            α, β, backend, allocator
+        )
+    end
 
     codA, domA = codomainind(A), domainind(A)
     codB, domB = codomainind(B), domainind(B)

src/tensors/diagonal.jl

@@ -280,7 +280,7 @@ end
 # ----------------
 function TO.tensoradd_type(TC, A::DiagonalTensorMap, ::Index2Tuple{1, 1}, ::Bool)
     M = similarstoragetype(A, TC)
-    return DiagonalTensorMap{TC, spacetype(A), M}
+    return DiagonalTensorMap{scalartype(M), spacetype(A), M}
 end
 
 function TO.tensorcontract_type(

src/tensors/indexmanipulations.jl

@@ -17,6 +17,8 @@ for (operation, manipulation) in (
         $promote_op(::Type{T}, ::Type{I}) where {T <: Number, I <: Sector} =
             sectorscalartype(I) <: Integer ? T :
             sectorscalartype(I) <: Real ? float(T) : complex(T)
+        $promote_op(::Type{TA}, ::Type{I}) where {TA <: DenseVector, I <: Sector} =
+            similarstoragetype(TA, $promote_op(eltype(TA), I))
         # TODO: currently the manipulations all use sectorscalartype, change to:
         # $manipulation_scalartype(I) <: Integer ? T :
         # $manipulation_scalartype(I) <: Real ? float(T) : complex(T)
@@ -342,11 +344,11 @@ See also [`insertrightunit`](@ref insertrightunit(::AbstractTensorMap, ::Val{i})
 """
 function insertleftunit(
         t::AbstractTensorMap, ::Val{i} = Val(numind(t) + 1);
-        copy::Bool = false, conj::Bool = false, dual::Bool = false
+        copy::Bool = false, conj::Bool = false, dual::Bool = false,
     ) where {i}
     W = insertleftunit(space(t), Val(i); conj, dual)
     if t isa TensorMap
-        return TensorMap{scalartype(t)}(copy ? Base.copy(t.data) : t.data, W)
+        return TensorMapWithStorage{scalartype(t), storagetype(t)}(copy ? Base.copy(t.data) : t.data, W)
     else
         tdst = similar(t, W)
         for (c, b) in blocks(t)
@@ -371,11 +373,11 @@ See also [`insertleftunit`](@ref insertleftunit(::AbstractTensorMap, ::Val{i}) w
 """
 function insertrightunit(
         t::AbstractTensorMap, ::Val{i} = Val(numind(t));
-        copy::Bool = false, conj::Bool = false, dual::Bool = false
+        copy::Bool = false, conj::Bool = false, dual::Bool = false,
     ) where {i}
     W = insertrightunit(space(t), Val(i); conj, dual)
     if t isa TensorMap
-        return TensorMap{scalartype(t)}(copy ? Base.copy(t.data) : t.data, W)
+        return TensorMapWithStorage{scalartype(t), storagetype(t)}(copy ? Base.copy(t.data) : t.data, W)
     else
         tdst = similar(t, W)
         for (c, b) in blocks(t)
@@ -400,7 +402,7 @@ and [`insertrightunit`](@ref insertrightunit(::AbstractTensorMap, ::Val{i}) wher
 function removeunit(t::AbstractTensorMap, ::Val{i}; copy::Bool = false) where {i}
     W = removeunit(space(t), Val(i))
     if t isa TensorMap
-        return TensorMap{scalartype(t)}(copy ? Base.copy(t.data) : t.data, W)
+        return TensorMapWithStorage{scalartype(t), storagetype(t)}(copy ? Base.copy(t.data) : t.data, W)
     else
         tdst = similar(t, W)
         for (c, b) in blocks(t)