Updates

Author: kshyatt

Project.toml

@@ -18,23 +18,24 @@ TupleTools = "9d95972d-f1c8-5527-a6e0-b4b365fa01f6"
 VectorInterface = "409d34a3-91d5-4945-b6ec-7529ddf182d8"
 
 [weakdeps]
-AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
+#AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 FiniteDifferences = "26cc04aa-876d-5657-8c51-4c34ba976000"
 
 [extensions]
-TensorKitAMDGPUExt = "AMDGPU"
+#TensorKitAMDGPUExt = "AMDGPU"
 TensorKitCUDAExt = ["CUDA", "cuTENSOR"]
 TensorKitChainRulesCoreExt = "ChainRulesCore"
 TensorKitFiniteDifferencesExt = "FiniteDifferences"
 
 [compat]
-AMDGPU = "2"
+#AMDGPU = "2"
 Adapt = "4"
 Aqua = "0.6, 0.7, 0.8"
 CUDA = "5"
+cuTENSOR = "2.2"
 ChainRulesCore = "1"
 ChainRulesTestUtils = "1"
 Combinatorics = "1"
@@ -54,12 +55,13 @@ TestExtras = "0.2,0.3"
 TupleTools = "1.1"
 VectorInterface = "0.4, 0.5"
 Zygote = "0.7"
-cuTENSOR = "2"
 julia = "1.10"
 
 [extras]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 ChainRulesTestUtils = "cdddcdb0-9152-4a09-a978-84456f9df70a"
 Combinatorics = "861a8166-3701-5b0c-9a16-15d98fcdc6aa"
@@ -71,4 +73,9 @@ TestExtras = "5ed8adda-3752-4e41-b88a-e8b09835ee3a"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["Adapt", "AMDGPU", "Aqua", "Combinatorics", "CUDA", "cuTENSOR", "LinearAlgebra", "TensorOperations", "Test", "TestExtras", "ChainRulesCore", "ChainRulesTestUtils", "FiniteDifferences", "Zygote"]
+test = ["Adapt", "Aqua", "Combinatorics", "CUDA", "cuTENSOR", "LinearAlgebra", "TensorOperations", "Test", "TestExtras", "ChainRulesCore", "ChainRulesTestUtils", "FiniteDifferences", "Zygote"]
+
+[sources]
+CUDA = {url = "https://github.com/JuliaGPU/CUDA.jl", rev = "master"}
+cuTENSOR = {url = "https://github.com/JuliaGPU/CUDA.jl", subdir="lib/cutensor", rev = "ksh/cutensor_bump"}
+MatrixAlgebraKit = {url = "https://github.com/QuantumKitHub/MatrixAlgebraKit.jl", rev = "ksh/tk"}

ext/TensorKitCUDAExt/TensorKitCUDAExt.jl

@@ -16,7 +16,7 @@ using Random
 
 include("cutensormap.jl")
 
-TensorKit.Factorizations.select_svd_algorithm(::CuTensorMap, ::TensorKit.Factorizations.SVD) = CUSOLVER_Jacobi()
+TensorKit.Factorizations.select_svd_algorithm(::CuTensorMap, ::TensorKit.Factorizations.SVD) = CUSOLVER_QRIteration()
 TensorKit.Factorizations.select_svd_algorithm(::CuTensorMap, ::TensorKit.Factorizations.SDD) = throw(ArgumentError("DivideAndConquer unavailable on CUDA")) 
 TensorKit.Factorizations.select_svd_algorithm(::CuTensorMap, alg::OFA) = throw(ArgumentError(lazy"Unknown algorithm $alg"))
 

test/cuda.jl

@@ -160,6 +160,7 @@ for V in spacelist
             @test LinearAlgebra.isdiag(D)
             @test LinearAlgebra.diag(D) == d
         end=#
+        #=
         @timedtestset "Permutations: test via inner product invariance" begin
             W = V1 ⊗ V2 ⊗ V3 ⊗ V4 ⊗ V5
             t = CUDA.rand(ComplexF64, W)
@@ -323,7 +324,7 @@ for V in spacelist
                     @test ad(t1' / t') ≈ ad(t1)' / ad(t)'
                 end
             end
-        end
+        end=#
         @timedtestset "Factorization" begin
             W = V1 ⊗ V2 ⊗ V3 ⊗ V4 ⊗ V5
             for T in (Float32, ComplexF64)
@@ -336,20 +337,14 @@ for V in spacelist
                                                        TensorKit.Polar(), TensorKit.SVD(),
                                                        )
                         Q, R = @constinferred leftorth(t, ((3, 4, 2), (1, 5)); alg=alg)
-                        QdQ = Q' * Q
-                        @test QdQ ≈ one(QdQ)
+                        @test isisometry(Q)
                         @test Q * R ≈ permute(t, ((3, 4, 2), (1, 5)))
-                        if alg isa Polar
-                            # @test isposdef(R) # not defined for CUDA
-                            @test_broken domain(R) == codomain(R) == space(t, 1)' ⊗ space(t, 5)'
-                        end
                     end
                     @testset "leftnull with $alg" for alg in
                                                       (TensorKit.QR(), TensorKit.SVD(),
                                                        )
                         N = @constinferred leftnull(t, ((3, 4, 2), (1, 5)); alg=alg)
-                        NdN = N' * N
-                        @test NdN ≈ one(NdN)
+                        @test isisometry(N)
                         @test norm(N' * permute(t, ((3, 4, 2), (1, 5)))) <
                               100 * eps(norm(t))
                     end
@@ -360,29 +355,21 @@ for V in spacelist
                                                         )
                         # cusolver SVD requires m >= n for some reason
                         L, Q = @constinferred rightorth(t, ((3, 4), (2, 1, 5)); alg=alg)
-                        QQd = Q * Q'
-                        @test QQd ≈ one(QQd)
+                        @test isisometry(Q; side=:right)
                         @test L * Q ≈ permute(t, ((3, 4), (2, 1, 5)))
-                        if alg isa Polar
-                            # @test isposdef(L) # not defined for CUDA
-                            @test_broken domain(L) == codomain(L) == space(t, 3) ⊗ space(t, 4)
-                        end
                     end
                     @testset "rightnull with $alg" for alg in
                                                        (TensorKit.LQ(), TensorKit.SVD(),
                                                         )
                         M = @constinferred rightnull(t, ((3, 4), (2, 1, 5)); alg=alg)
-                        MMd = M * M'
-                        @test MMd ≈ one(MMd)
+                        @test isisometry(M; side=:right)
                         @test norm(permute(t, ((3, 4), (2, 1, 5))) * M') <
                               100 * eps(norm(t))
                     end
                     @testset "tsvd with $alg" for alg in (TensorKit.SVD(),)
                         U, S, V = @constinferred tsvd(t, ((3, 4, 2), (1, 5)); alg=alg)
-                        UdU = U' * U
-                        @test UdU ≈ one(UdU)
-                        VVd = V * V'
-                        @test VVd ≈ one(VVd)
+                        @test isisometry(U)
+                        @test isisometry(V; side=:right)
                         t2 = permute(t, ((3, 4, 2), (1, 5)))
                         US = U * S
                         USV = US * V