More tests for CUDA + MPS

Project.toml

@@ -76,3 +76,8 @@ cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 
 [targets]
 test = ["Aqua", "Adapt", "CUDA", "cuTENSOR", "Pkg", "Test", "TestExtras", "Plots", "Combinatorics", "ParallelTestRunner", "TensorKitTensors"]
+
+[sources]
+TensorKit = {url = "https://github.com/QuantumKitHub/TensorKit.jl", rev = "ksh/cuda_tweaks"}
+BlockTensorKit = {url = "https://github.com/kshyatt/BlockTensorKit.jl", rev = "ksh/gpu"}
+

ext/MPSKitAdaptExt.jl

@@ -35,5 +35,7 @@ end
     MPSKit.JordanMPOTensor(space(W), adapt(to, W.A), adapt(to, W.B), adapt(to, W.C), adapt(to, W.D))
 @inline Adapt.adapt_structure(to, mpo::MPOHamiltonian) =
     MPOHamiltonian(map(x -> adapt(to, x), mpo.W))
+@inline Adapt.adapt_structure(to, ml::MPSKit.Multiline) = MPSKit.Multiline(map(adapt(to), ml.data))
+@inline Adapt.adapt_structure(to, pa::MPSKit.PeriodicArray) = MPSKit.PeriodicArray(map(adapt(to), pa.data))
 
 end

src/operators/abstractmpo.jl

@@ -99,7 +99,7 @@ Base.:\(α::Number, mpo::AbstractMPO) = scale(mpo, inv(α))
 
 function VectorInterface.scale(mpo::AbstractMPO, α::Number)
     T = VectorInterface.promote_scale(scalartype(mpo), scalartype(α))
-    dst = similar(mpo, T)
+    dst = similar(mpo, TensorKit.similarstoragetype(storagetype(mpo), T))
     return scale!(dst, mpo, α)
 end
 
@@ -152,7 +152,11 @@ Compute the mpo tensor that arises from multiplying MPOs.
 """
 function fuse_mul_mpo(O1, O2)
     TT = promote_type(scalartype(O1), scalartype(O2))
-    T = TensorKit.similarstoragetype(storagetype(O1), TT)
+    T = if O1 isa BraidingTensor
+        TensorKit.similarstoragetype(storagetype(O2), TT)
+    else
+        TensorKit.similarstoragetype(storagetype(O1), TT)
+    end
     F_left = fuser(T, left_virtualspace(O2), left_virtualspace(O1))
     F_right = fuser(T, right_virtualspace(O2), right_virtualspace(O1))
     return _fuse_mpo_mpo(O1, O2, F_left, F_right)

src/operators/jordanmpotensor.jl

@@ -135,6 +135,10 @@ function jordanmpotensortype(::Type{O}) where {O <: AbstractTensorMap}
 end
 Base.similar(W::JordanMPOTensor, ::Type{TorA}) where {TorA} =
     jordanmpotensortype(spacetype(W), TorA)(undef, space(W))
+Base.similar(W::JordanMPOTensor, V::VectorSpace) =
+    jordanmpotensortype(spacetype(V), storagetype(W))(undef, V)
+Base.similar(W::JordanMPOTensor, V::VectorSpace, ::Type{TorA}) where {TorA} =
+    jordanmpotensortype(spacetype(V), TorA)(undef, V)
 
 # Properties
 # ----------
@@ -205,7 +209,8 @@ end
 for f in (:real, :complex)
     @eval function Base.$f(W::JordanMPOTensor)
         E = $f(scalartype(W))
-        W′ = JordanMPOTensor{E}(undef, space(W))
+        TE = TensorKit.similarstoragetype(TensorKit.storagetype(W), E)
+        W′ = similar(W, TE)
         for (I, v) in nonzero_pairs(W.A)
             W′.A[I] = $f(v)
         end
@@ -242,7 +247,7 @@ end
     elseif 1 < i < size(W, 1) && 1 < j < size(W, 4)
         return W.A[i - 1, 1, 1, j - 1]
     else
-        return zeros(scalartype(W), eachspace(W)[i, 1, 1, j])
+        return zeros(storagetype(W), eachspace(W)[i, 1, 1, j])
     end
 end
 
@@ -357,7 +362,7 @@ function add_physical_charge(O::JordanMPOTensor, charge::Sector)
     Vdst = left_virtualspace(O) ⊗
         fuse(physicalspace(O), auxspace) ←
         fuse(physicalspace(O), auxspace) ⊗ right_virtualspace(O)
-    Odst = JordanMPOTensor{scalartype(O)}(undef, Vdst)
+    Odst = similar(O, Vdst)
     for (I, v) in nonzero_pairs(O)
         Odst[I] = add_physical_charge(v, charge)
     end

src/operators/mpo.jl

@@ -62,7 +62,7 @@ eachsite(mpo::InfiniteMPO) = PeriodicArray(eachindex(mpo))
 function Base.similar(mpo::MPO{<:MPOTensor}, ::Type{O}, L::Int) where {O <: MPOTensor}
     return MPO(similar(parent(mpo), O, L))
 end
-function Base.similar(mpo::MPO, ::Type{T}) where {T <: Number}
+function Base.similar(mpo::MPO, ::Type{T}) where {T <: Union{Number, DenseVector}}
     return MPO(similar.(parent(mpo), T))
 end
 

src/operators/mpohamiltonian.jl

@@ -503,10 +503,8 @@ function InfiniteMPOHamiltonian(lattice′::AbstractArray{<:VectorSpace}, local_
     end
 
     # partial sort by interaction range
-    local_mpos = sort!(
-        map(Base.Fix1(instantiate_operator, lattice), collect(local_operators));
-        by = x -> length(x[1])
-    )
+    unsorted_mpos = map(Base.Fix1(instantiate_operator, lattice), [local_operators...])
+    local_mpos = sort!(unsorted_mpos; by = x -> length(x[1]))
 
     for (sites, local_mpo) in local_mpos
         local key_R # trick to define key_R before the first iteration
@@ -662,6 +660,8 @@ function isemptylevel(H::InfiniteMPOHamiltonian, i::Int)
 end
 
 function Base.convert(::Type{TensorMap}, H::FiniteMPOHamiltonian)
+    # TODO this FORCES the output tensor to have Vector storage
+    # in an uncontrollable way, should be fixed
     L = removeunit(H[1], 1)
     R = removeunit(H[end], 4)
     M = Tuple(H[2:(end - 1)])
@@ -701,8 +701,8 @@ end
 function Base.similar(H::MPOHamiltonian, ::Type{O}, L::Int) where {O <: MPOTensor}
     return MPOHamiltonian(similar(parent(H), O, L))
 end
-function Base.similar(H::MPOHamiltonian, ::Type{T}) where {T <: Number}
-    return MPOHamiltonian(similar.(parent(H), T))
+function Base.similar(H::MPOHamiltonian, ::Type{TorA}) where {TorA <: Union{Number, DenseVector}}
+    return MPOHamiltonian(similar.(parent(H), TorA))
 end
 
 # Linear Algebra
@@ -727,7 +727,7 @@ function Base.:+(
             ⊞(Vtriv, right_virtualspace(A), Vtriv)
         V = Vleft ⊗ physicalspace(A) ← physicalspace(A) ⊗ Vright
 
-        H[i] = eltype(H)(V, A, B, C, D)
+        H[i] = JordanMPOTensor(V, A, B, C, D)
     end
     return FiniteMPOHamiltonian(H)
 end
@@ -749,7 +749,7 @@ function Base.:+(
         Vright = ⊞(Vtriv, right_virtualspace(A), Vtriv)
         V = Vleft ⊗ physicalspace(A) ← physicalspace(A) ⊗ Vright
 
-        H[i] = eltype(H)(V, A, B, C, D)
+        H[i] = JordanMPOTensor(V, A, B, C, D)
     end
     return InfiniteMPOHamiltonian(H)
 end
@@ -855,7 +855,13 @@ end
 function Base.:*(H::FiniteMPOHamiltonian{<:MPOTensor}, x::AbstractTensorMap)
     @assert length(H) > 1
     @assert numout(x) == length(H)
-    L = removeunit(H[1], 1)
+    if H[1] isa BraidingTensor
+        L′ = removeunit(H[1], 1)
+        L = similar(L′, TensorKit.storagetype(H))
+        copy!(L, L′)
+    else
+        L = removeunit(H[1], 1)
+    end
     M = Tuple(H[2:(end - 1)])
     R = removeunit(H[end], 4)
     return TensorMap(_apply_finitempo(x, L, M, R))

src/operators/ortho.jl

@@ -11,7 +11,14 @@ function left_canonicalize!(
     d = sqrt(dim(P))
 
     # orthogonalize second column against first
-    WI = removeunit(W[1, 1, 1, 1], 1)
+    Wi = W[1, 1, 1, 1]
+    if Wi isa BraidingTensor
+        Wi′ = removeunit(Wi, 1)
+        WI = similar(Wi′, TensorKit.storagetype(H))
+        copy!(WI, Wi′)
+    else
+        WI = removeunit(Wi, 1)
+    end
     @plansor t[l; r] := conj(WI[p; p' l]) * W.C[p; p' r]
     # TODO: the following is currently broken due to a TensorKit bug
     # @plansor C′[p; p' r] := W.C[p; p' r] - WI[p; p' l] * t[l; r]
@@ -99,7 +106,14 @@ function right_canonicalize!(
     d = sqrt(dim(P))
 
     # orthogonalize second row against last
-    WI = removeunit(W[end, 1, 1, end], 4)
+    Wi = W[end, 1, 1, end]
+    if Wi isa BraidingTensor
+        Wi′ = removeunit(Wi, 4)
+        WI = similar(Wi′, TensorKit.storagetype(H))
+        copy!(WI, Wi′)
+    else
+        WI = removeunit(Wi, 4)
+    end
     @plansor t[l; r] := conj(WI[r p; p']) * W.B[l p; p']
     # TODO: the following is currently broken due to a TensorKit bug
     # @plansor B′[l p; p'] := W.B[l p; p'] - WI[r p; p'] * t[l; r]
@@ -124,7 +138,8 @@ function right_canonicalize!(
         end
         H[i] = JordanMPOTensor(V ⊗ P ← domain(W), Q1, Q2, W.C, W.D)
     else
-        tmp = transpose(cat(insertleftunit(B′, 4), W.A; dims = 4), ((1,), (3, 4, 2)))
+        B′′ = insertleftunit(B′, 4)
+        tmp = transpose(cat(B′′, W.A; dims = 4), ((1,), (3, 4, 2)))
         R, Q = right_orth!(tmp; alg)
         if dim(R) == 0
             V = _rightunit ⊞ _rightunit

test/cuda/operators.jl

@@ -5,11 +5,15 @@ using MPSKit: GeometryStyle, FiniteChainStyle, InfiniteChainStyle, OperatorStyle
 using TensorKit
 using MatrixAlgebraKit
 using TensorKit: ℙ, tensormaptype, TensorMapWithStorage
-using Adapt, CUDA, cuTENSOR
+using Adapt, CUDA, cuTENSOR, CUDA.CUBLAS
 
 # TODO revisit this once https://github.com/QuantumKitHub/MatrixAlgebraKit.jl/issues/176
 # is resolved
 MPSKit.Defaults.alg_svd() = CUSOLVER_QRIteration()
+MPSKit.Defaults.alg_lq() = LQViaTransposedQR(CUSOLVER_HouseholderQR())
+
+pspaces = (ℙ^4, Rep[U₁](0 => 2), Rep[SU₂](1 => 1))
+vspaces = (ℙ^10, Rep[U₁]((0 => 20)), Rep[SU₂](1 // 2 => 10, 3 // 2 => 5, 5 // 2 => 1))
 
 @testset "CuFiniteMPO" for V in (ℂ^2, U1Space(0 => 1, 1 => 1))
     # start from random operators
@@ -87,3 +91,171 @@ MPSKit.Defaults.alg_svd() = CUSOLVER_QRIteration()
 
     @test dot(mpomps₁, mpomps₁) ≈ dot(mpo₁, mpo₁)
 end
+
+@testset "Finite CuMPOHamiltonian T ($T), V ($(spacetype(V)))" for T in (Float64, ComplexF64), V in (ℂ^2, U1Space(-1 => 1, 0 => 1, 1 => 1))
+    L = 3
+    lattice = fill(V, L)
+    O₁ = randn(T, V, V)
+    O₁ += O₁'
+    E = id(CuVector{T, CUDA.DeviceMemory}, domain(O₁))
+    O₂ = randn(T, V^2 ← V^2)
+    O₂ += O₂'
+
+    dO₁ = adapt(CuVector{T, CUDA.DeviceMemory}, O₁)
+    dO₂ = adapt(CuVector{T, CUDA.DeviceMemory}, O₂)
+    H1 = adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(lattice, i => O₁ for i in 1:L))
+    H2 = adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(lattice, (i, i + 1) => O₂ for i in 1:(L - 1)))
+    H3 = adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(lattice, 1 => O₁, (2, 3) => O₂, (1, 3) => O₂))
+    @test TensorKit.storagetype(H1) == CuVector{T, CUDA.DeviceMemory}
+    @test TensorKit.storagetype(H2) == CuVector{T, CUDA.DeviceMemory}
+    @test TensorKit.storagetype(H3) == CuVector{T, CUDA.DeviceMemory}
+
+    @test scalartype(H1) == scalartype(H2) == scalartype(H3) == T
+    if !(T <: Complex)
+        for H in (H1, H2, H3)
+            Hc = @constinferred complex(H)
+            @test scalartype(Hc) == complex(T)
+            @test TensorKit.storagetype(Hc) == CuVector{complex(T), CUDA.DeviceMemory}
+        end
+    end
+
+    # check if constructor works by converting back to tensormap
+    H1_tm = convert(TensorMap, H1)
+    @test TensorKit.storagetype(H1_tm) == CuVector{T, CUDA.DeviceMemory}
+    operators = vcat(fill(E, L - 1), dO₁)
+    @test H1_tm ≈ mapreduce(+, 1:L) do i
+        return reduce(⊗, circshift(operators, i))
+    end
+    operators = vcat(fill(E, L - 2), dO₂)
+    @test convert(TensorMap, H2) ≈ mapreduce(+, 1:(L - 1)) do i
+        return reduce(⊗, circshift(operators, i))
+    end
+    @test convert(TensorMap, H3) ≈
+        dO₁ ⊗ E ⊗ E + E ⊗ dO₂ + permute(dO₂ ⊗ E, ((1, 3, 2), (4, 6, 5)))
+
+    # check if adding terms on the same site works
+    single_terms = Iterators.flatten(Iterators.repeated((i => O₁ / 2 for i in 1:L), 2))
+    H4 = adapt(CuArray, FiniteMPOHamiltonian(lattice, single_terms))
+    @test TensorKit.storagetype(H4) == CuVector{T, CUDA.DeviceMemory}
+    @test H4 ≈ H1 atol = 1.0e-6
+    double_terms = Iterators.flatten(
+        Iterators.repeated(((i, i + 1) => O₂ / 2 for i in 1:(L - 1)), 2)
+    )
+    H5 = adapt(CuArray, FiniteMPOHamiltonian(lattice, double_terms))
+    @test TensorKit.storagetype(H5) == CuVector{T, CUDA.DeviceMemory}
+    @test H5 ≈ H2 atol = 1.0e-6
+
+    # test linear algebra
+    @test H1 ≈
+        adapt(CuArray, FiniteMPOHamiltonian(lattice, 1 => O₁)) +
+        adapt(CuArray, FiniteMPOHamiltonian(lattice, 2 => O₁)) +
+        adapt(CuArray, FiniteMPOHamiltonian(lattice, 3 => O₁))
+    @test TensorKit.storagetype(H1) == CuVector{T, CUDA.DeviceMemory}
+
+    H1′a = 0.8 * H1
+    @test TensorKit.storagetype(H1′a) == CuVector{T, CUDA.DeviceMemory}
+    H1′b = 0.2 * H1
+    @test TensorKit.storagetype(H1′b) == CuVector{T, CUDA.DeviceMemory}
+    H1′ = H1′a + H1′b
+    @test TensorKit.storagetype(H1′) == CuVector{T, CUDA.DeviceMemory}
+    @test H1′ ≈ H1 atol = 1.0e-6
+    @test convert(TensorMap, H1 + H2) ≈ convert(TensorMap, H1) + convert(TensorMap, H2) atol = 1.0e-6
+    H1_trunc = changebonds(H1, SvdCut(; trscheme = truncrank(0)))
+    @test H1_trunc ≈ H1
+    @test all(left_virtualspace(H1_trunc) .== left_virtualspace(H1))
+
+    # test dot and application
+    state = adapt(CuVector{T, CUDA.DeviceMemory}, rand(T, prod(lattice)))
+    @test TensorKit.storagetype(state) == CuVector{T, CUDA.DeviceMemory}
+    mps = adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPS(state))
+    @test TensorKit.storagetype(mps) == CuVector{T, CUDA.DeviceMemory}
+
+    # TODO fix me!
+    #=@test TensorKit.storagetype(H1) == CuVector{T, CUDA.DeviceMemory}
+    H1mps = H1 * mps
+    @test TensorKit.storagetype(H1mps) == CuVector{T, CUDA.DeviceMemory}
+    H1tmstate = H1_tm * state 
+    @test TensorKit.storagetype(H1tmstate) == CuVector{T, CUDA.DeviceMemory}
+    H1mpstm = convert(TensorMap, H1mps)
+    @test TensorKit.storagetype(H1mpstm) == CuVector{T, CUDA.DeviceMemory}
+    @test H1mpstm ≈ H1tmstate
+    @test H1 * state ≈ H1tmstate
+    @test dot(mps, H2, mps) ≈ dot(mps, H2 * mps)=#
+
+    # test constructor from dictionary with mixed linear and Cartesian lattice indices as keys
+    grid = square = fill(V, 3, 3)
+
+    local_operators = Dict((I,) => O₁ for I in eachindex(grid))
+    I_vertical = CartesianIndex(1, 0)
+    vertical_operators = Dict(
+        (I, I + I_vertical) => O₂ for I in eachindex(IndexCartesian(), square) if I[1] < size(square, 1)
+    )
+    I_horizontal = CartesianIndex(0, 1)
+    horizontal_operators = Dict(
+        (I, I + I_horizontal) => O₂ for I in eachindex(IndexCartesian(), square) if I[2] < size(square, 1)
+    )
+    operators = merge(local_operators, vertical_operators, horizontal_operators)
+    H4 = adapt(CuArray, FiniteMPOHamiltonian(grid, operators))
+    @test TensorKit.storagetype(H4) == CuVector{T, CUDA.DeviceMemory}
+
+    @test H4 ≈
+        adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(grid, local_operators)) +
+        adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(grid, vertical_operators)) +
+        adapt(CuVector{T, CUDA.DeviceMemory}, FiniteMPOHamiltonian(grid, horizontal_operators)) atol = 1.0e-4
+    @test TensorKit.storagetype(H4) == CuVector{T, CUDA.DeviceMemory}
+
+    H4′ = H4 / 3 + 2H4 / 3
+    @test TensorKit.storagetype(H4′) == CuVector{T, CUDA.DeviceMemory}
+    #=H5 = changebonds(H4′, SvdCut(; trscheme = trunctol(; atol = 1.0e-12)))
+    @test TensorKit.storagetype(H5) == CuVector{T, CUDA.DeviceMemory}
+    psi = adapt(CuArray, FiniteMPS(physicalspace(H5), V ⊕ rightunitspace(V)))
+    @test expectation_value(psi, H4) ≈ expectation_value(psi, H5)=# # TODO
+end
+
+@testset "CuInfiniteMPOHamiltonian $(sectortype(pspace))" for (pspace, Dspace) in zip(pspaces, vspaces)
+    # generate a 1-2-3 body interaction
+    operators = ntuple(3) do i
+        O = rand(ComplexF64, pspace^i, pspace^i)
+        return O += O'
+    end
+
+    H1 = adapt(CuVector{ComplexF64, CUDA.DeviceMemory}, InfiniteMPOHamiltonian(operators[1]))
+    H2 = adapt(CuVector{ComplexF64, CUDA.DeviceMemory}, InfiniteMPOHamiltonian(operators[2]))
+    H3 = adapt(CuVector{ComplexF64, CUDA.DeviceMemory}, repeat(InfiniteMPOHamiltonian(operators[3]), 2))
+
+    @test TensorKit.storagetype(H1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test TensorKit.storagetype(H2) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test TensorKit.storagetype(H3) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    # make a teststate to measure expectation values for
+    ψ1 = adapt(CuVector{ComplexF64, CUDA.DeviceMemory}, InfiniteMPS([pspace], [Dspace]))
+    ψ2 = adapt(CuVector{ComplexF64, CUDA.DeviceMemory}, InfiniteMPS([pspace, pspace], [Dspace, Dspace]))
+    @test TensorKit.storagetype(ψ1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test TensorKit.storagetype(ψ2) == CuVector{ComplexF64, CUDA.DeviceMemory}
+
+    e1 = expectation_value(ψ1, H1)
+    e2 = expectation_value(ψ1, H2)
+
+    H1 = 2 * H1
+    @test TensorKit.storagetype(H1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    H1 -= [1]
+    @test TensorKit.storagetype(H1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test e1 * 2 - 1 ≈ expectation_value(ψ1, H1) atol = 1.0e-10
+
+    H1 = H1 + H2
+    @test TensorKit.storagetype(H1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test e1 * 2 + e2 - 1 ≈ expectation_value(ψ1, H1) atol = 1.0e-10
+
+    H1 = repeat(H1, 2)
+    @test TensorKit.storagetype(H1) == CuVector{ComplexF64, CUDA.DeviceMemory}
+
+    e1 = expectation_value(ψ2, H1)
+    e3 = expectation_value(ψ2, H3)
+
+    @test e1 + e3 ≈ expectation_value(ψ2, H1 + H3) atol = 1.0e-10
+
+    H4 = H1 + H3
+    @test TensorKit.storagetype(H4) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    h4 = H4 * H4
+    @test TensorKit.storagetype(h4) == CuVector{ComplexF64, CUDA.DeviceMemory}
+    @test real(expectation_value(ψ2, H4)) >= 0
+end