Support to Unified memory arrays

[albertomercurio]

I'm trying the example in the README on CSR matrices where the underlying pointers are in unified memory.

It works on standard size, but I get out of memory easily as I were running of device memory

ERROR: LoadError: Out of GPU memory
Effective GPU memory usage: 6.37% (1.498 GiB/23.508 GiB)
Memory pool usage: 182.493 MiB (192.000 MiB reserved)

Stacktrace:
  [1] throw_api_error(res::CUDSS.cudssStatus_t)
    @ CUDSS ~/.julia/packages/CUDSS/LDGbz/src/error.jl:35
  [2] check
    @ ~/.julia/packages/CUDSS/LDGbz/src/error.jl:48 [inlined]
  [3] cudssExecute
    @ ~/.julia/packages/GPUToolbox/JLBB1/src/ccalls.jl:33 [inlined]
  [4] cudss(phase::String, solver::CudssSolver{…}, X::CudssMatrix{…}, B::CudssMatrix{…}; asynchronous::Bool)
    @ CUDSS ~/.julia/packages/CUDSS/LDGbz/src/interfaces.jl:504
  [5] cudss
    @ ~/.julia/packages/CUDSS/LDGbz/src/interfaces.jl:501 [inlined]
  [6] lu!(solver::CudssSolver{ComplexF64, Int32}, A::CuSparseMatrixCSR{ComplexF64, Int32}; check::Bool)
    @ CUDSS ~/.julia/packages/CUDSS/LDGbz/src/generic.jl:37
  [7] lu!
    @ ~/.julia/packages/CUDSS/LDGbz/src/generic.jl:31 [inlined]

[amontoison]

How do you setup the unify memory? Do you a LocalPreferences.toml to force unify memory everywhere?

[albertomercurio]

using CUDA, CUDA.CUSPARSE
using CUDSS
using SparseArrays, LinearAlgebra
using SparseMatricesCSR

T = Float64
n = 100
A_cpu = sprand(T, n, n, 0.05) + I
x_cpu = zeros(T, n)
b_cpu = rand(T, n)

A_cpu_csr = SparseMatrixCSR(A_cpu)

rowPtr = CuVector{eltype(A_cpu_csr.rowptr), CUDA.UnifiedMemory}(A_cpu_csr.rowptr)
colVal = CuVector{eltype(A_cpu_csr.colval), CUDA.UnifiedMemory}(A_cpu_csr.colval)
nzVal  = CuVector{eltype(A_cpu_csr.nzval), CUDA.UnifiedMemory}(T.(A_cpu_csr.nzval))

A_gpu = CuSparseMatrixCSR(rowPtr, colVal, nzVal, (size(A_cpu_csr, 1), size(A_cpu_csr, 2)))
x_gpu = T.(cu(x_cpu; unified=true))
b_gpu = T.(cu(b_cpu; unified=true))

solver = CudssSolver(A_gpu, "G", 'F')

cudss("analysis", solver, x_gpu, b_gpu)
cudss("factorization", solver, x_gpu, b_gpu)
cudss("solve", solver, x_gpu, b_gpu)

I don't have any LocalPreferences.toml

[amontoison]

@albertomercurio

amontoison@kkt:~$ julia
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.11.7 (2025-09-08)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> using CUDA, CUDA.CUSPARSE

julia> using CUDSS

julia> using SparseArrays, LinearAlgebra

julia> using SparseMatricesCSR
 │ Package SparseMatricesCSR not found, but a package named SparseMatricesCSR is available from a
 │ registry. 
 │ Install package?
 │   (@v1.11) pkg> add SparseMatricesCSR 
 └ (y/n/o) [y]: y
   Resolving package versions...
   Installed SparseMatricesCSR ─ v0.6.9
    Updating `~/.julia/environments/v1.11/Project.toml`
  [a0a7dd2c] + SparseMatricesCSR v0.6.9
    Updating `~/.julia/environments/v1.11/Manifest.toml`
  [a0a7dd2c] + SparseMatricesCSR v0.6.9
  [4607b0f0] + SuiteSparse
Precompiling SparseMatricesCSR...
  1 dependency successfully precompiled in 1 seconds. 4 already precompiled.
Precompiling SparseMatricesCSRExt...
  1 dependency successfully precompiled in 6 seconds. 107 already precompiled.

julia> T = Float64
Float64

julia> n = 100
100

julia> A_cpu = sprand(T, n, n, 0.05) + I
100×100 SparseMatrixCSC{Float64, Int64} with 595 stored entries:
⎡⠑⢄⠀⠢⠀⠈⠀⠀⠀⠁⠀⠀⠀⡀⠀⠀⠀⠀⠂⡠⠀⠀⢀⠄⠀⠀⡈⠀⠠⠀⠀⠀⠐⠀⠀⠀⡀⠀⠠⢁⠀⠨⡀⠀⠄⠀⠀⠆⠀⠀⎤
⎢⠀⠀⠑⣄⠀⠀⠀⢀⠀⠀⡀⠐⠀⠀⠀⠐⠀⠄⠄⠀⠀⠀⠐⠄⠀⠈⢀⠀⠀⠀⠈⠈⠀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠠⠀⣀⠀⠀⠁⠀⠦⎥
⎢⠠⠢⡀⠀⠑⢄⠒⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠁⠁⠀⠠⠀⠓⠀⠀⠀⠂⠀⠁⠠⠀⠀⠠⠀⠀⠀⠀⠀⠀⠄⠀⡀⠄⠀⠀⡀⠠⠀⠀⠑⎥
⎢⡀⠀⠂⠀⠀⠀⠑⢔⠐⠀⠈⠀⠀⠈⠀⠀⠀⠠⠀⠀⠠⠀⠀⠈⠀⢠⠀⠀⠀⡀⠀⠈⠀⠀⠀⠈⠀⠈⠀⠁⠀⠀⠀⠀⠀⡀⠂⠀⠀⠀⎥
⎢⠀⠀⠁⠀⠐⡀⠀⠠⠑⢄⠀⠀⠄⠀⠀⠀⠀⠀⠀⠀⠀⠌⠀⠀⠠⠀⠂⠐⠀⠀⠀⠀⠀⢀⢀⡨⡤⠀⠈⠂⠀⡀⠀⢀⢨⠀⠈⠀⠀⠀⎥
⎢⠠⠠⠀⠀⠀⠈⠀⡀⠀⠑⠑⢄⠀⠀⠠⠀⠀⠠⠀⠀⠀⠀⠀⠀⠀⠁⠀⠀⠀⠠⠀⠀⠀⠐⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠁⠀⠀⠀⠀⎥
⎢⠠⠁⠀⠀⠠⠀⠀⠀⠀⠀⠀⠁⠓⢄⠀⠀⠀⠀⠀⡀⠀⠂⠀⠀⠀⠄⠀⠠⠀⠀⠀⠀⡀⠠⠠⠀⠀⠠⠀⠈⠀⠐⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠈⠀⠀⠐⠀⠀⠈⠀⠀⠀⠀⠀⠑⢄⠀⠀⢀⠀⠁⠀⠀⠀⠠⠀⠄⡀⠀⠈⠀⠀⠄⠀⠀⠂⠀⠂⢀⡀⠠⠀⠈⢀⠀⢀⡀⠀⠀⢀⎥
⎢⠀⡀⠀⠀⢄⠀⠀⠀⠄⠀⠄⠀⠂⠀⠁⡐⠑⢄⠐⠀⠀⠁⠀⠀⠀⠀⠄⡀⠀⠀⠀⠁⠀⠂⠀⠁⠀⢀⠀⠀⠀⠀⡀⠀⢀⠀⠀⠀⠀⢀⎥
⎢⠀⢀⠄⣀⠀⡀⠠⠀⠀⠀⠀⠀⠀⢀⠠⠠⠀⠁⠑⢄⠀⠀⠀⠀⠀⠐⠂⠀⠀⠀⠀⠀⠀⠀⠀⠀⡀⠄⠀⠀⠀⠀⠀⠀⠀⠄⠀⠠⠀⠀⎥
⎢⠐⠢⠀⠄⠀⠀⠠⠀⠀⠀⠀⠁⠁⠂⠇⠀⠀⠀⠀⠀⢑⢌⠀⠀⠀⡈⠁⠀⠁⠀⠀⠄⡀⠂⠁⠀⠀⠀⠀⠀⢐⠀⠁⠀⠈⠆⠀⠀⠠⠀⎥
⎢⠁⠁⠐⠀⠂⠀⠀⡀⠀⠀⠀⠀⠢⠈⠐⠁⠀⠀⠀⠀⠀⠈⡑⢄⠀⠠⠀⠀⠀⠀⠀⠀⠀⠀⠁⠠⠀⠈⠕⠀⠀⠀⠀⠀⠀⡐⠀⠀⠀⠀⎥
⎢⠔⠀⡀⠀⠀⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠆⠠⠀⠀⠀⠀⠀⠀⡀⠑⢄⡀⠀⠀⠀⠂⠀⠐⢂⠀⠀⠀⢀⠀⠀⠀⢀⠑⠀⠀⠀⡀⠀⠀⠈⎥
⎢⠀⠀⠀⡀⠁⠀⠀⠀⠀⠀⢀⠀⠀⠈⠀⠀⠀⠀⠀⠀⢀⠀⠐⠀⠀⠀⠱⢄⠀⠀⠀⡀⠠⠈⠄⠀⠉⠐⠀⠀⠡⠊⠄⠈⠀⠀⠀⡁⠀⠀⎥
⎢⠀⡀⠀⠀⠐⠀⢀⠄⠀⠀⠀⠈⠀⠀⢀⠀⠂⠐⠀⡀⠀⠀⠁⠀⠀⢈⠠⠀⠑⣄⡀⠂⠅⠁⠁⢂⠀⠀⠀⠀⠀⠀⠐⠀⠀⠈⠀⠅⠁⠀⎥
⎢⠀⠀⠅⠀⠊⠀⠁⠀⠐⡀⠀⠀⠀⠀⠁⠈⢠⠀⠀⠠⠀⠐⠀⠀⠀⠀⠂⠀⠂⠐⠑⢅⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡈⠐⡀⠂⠀⠀⠀⎥
⎢⠀⠂⠀⠀⠀⠀⢀⠠⠀⠀⠀⠀⠀⠀⡀⠀⠀⠀⠀⠀⠠⠈⠈⠀⠊⠀⠀⠀⠀⠀⠀⠠⠑⢄⡀⠀⠀⠀⠀⠂⠀⠀⠀⢈⢀⠀⠀⠀⠄⠀⎥
⎢⠀⠂⠠⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠄⠀⠀⠀⠀⠀⡠⠀⠂⠀⠀⠀⢠⠁⠀⠀⠐⠂⠑⢆⠀⠀⠠⠀⠐⠀⠀⠀⠀⠠⠄⡀⠀⠀⎥
⎢⠠⠀⠀⠀⠀⠀⠈⠀⠀⠀⠀⠃⠐⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⠁⠄⠀⠀⠂⠀⠀⠀⠀⠀⠐⠀⠀⠕⢔⠀⡀⠀⠀⠀⠀⠀⠀⠀⠤⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⢀⠀⠄⠂⠀⢀⠀⢀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠀⠄⠀⠈⠀⠀⠀⠈⠀⠀⠀⠁⠑⢄⠀⠀⠂⠀⠀⠀⠀⠐⠀⠀⎥
⎢⡀⠀⠀⠀⠀⠀⠂⠀⠄⠈⠈⠀⠀⠀⠀⠂⠀⠀⠀⠀⠀⠀⠀⠀⠀⢘⠐⡀⠀⠀⠀⡁⠀⠀⠀⠁⠀⠄⠀⠀⠑⢄⠀⠐⠈⠀⠀⠀⠀⠁⎥
⎢⠀⠠⠀⠠⠀⠀⠀⠀⠀⠀⠀⠂⡀⠀⠀⠀⠂⢈⠀⠬⠀⠀⠑⠩⠀⠁⢀⠀⠀⠀⠰⡀⠀⠆⠀⠀⠀⠀⠐⡁⡀⠀⠑⢄⠂⠀⠀⠂⠄⠀⎥
⎢⠀⠆⠈⠀⠀⠀⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠀⠠⠀⠀⠀⠀⠀⠐⠀⠀⠀⠀⠀⠀⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡁⠀⠑⢄⢠⠄⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠄⠀⠀⠁⠈⠀⠀⠀⠀⢀⠄⡀⠀⠀⠀⠀⠀⠀⠀⢀⠐⠂⠈⠀⢁⠈⠈⠀⠀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡑⢄⡂⠀⎥
⎣⠀⠀⠀⠀⠀⡀⠊⡐⠀⠀⠀⠐⠀⢀⠀⠄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠆⠀⠀⠀⠀⠂⠀⠀⠐⠀⠀⠀⠀⠒⠠⠠⠀⠀⠊⠀⢠⠀⠠⠑⢄⎦

julia> x_cpu = zeros(T, n)
100-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 ⋮
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0

julia> b_cpu = rand(T, n)
100-element Vector{Float64}:
 0.6885662177311618
 0.19811875688587643
 0.34636657482973554
 0.508833856023195
 0.5786626733451149
 0.35277419578547353
 0.5900979754123136
 0.6028531490511718
 0.5654493189073513
 0.9753131371263993
 0.6991729644793807
 0.31716587183775136
 0.17102549549212198
 0.2767346240706057
 0.35442358111465433
 0.3042399282644813
 0.15991305416061685
 0.47176723369488827
 0.5561810153780739
 0.9483237082704531
 0.5645405082597065
 0.5958035725615859
 0.07342114388821663
 0.7556286483206618
 0.48321383462135004
 ⋮
 0.9282295302697778
 0.9285739010256263
 0.5164573886615473
 0.7207266613917147
 0.09463237057034557
 0.5399411888884054
 0.9454120741394579
 0.12503908501091043
 0.7671582177436156
 0.24666797761117398
 0.9779777813013302
 0.935341667839963
 0.824032172413455
 0.4694422991543259
 0.6423708873506008
 0.916535800400597
 0.6219422948946208
 0.2335545437140737
 0.20615165100331556
 0.4054636750686972
 0.8605594192574985
 0.6230349679552156
 0.2793578208255779
 0.42031103140836434
 0.09211910602918971

julia> A_cpu_csr = SparseMatrixCSR(A_cpu)
100×100 SparseMatrixCSR{1, Float64, Int64} with 595 stored entries:
  [1  ,   1]  =  1.0
  [1  ,  12]  =  0.941901
  [1  ,  19]  =  0.835937
  [1  ,  54]  =  0.232077
  [1  ,  79]  =  0.821251
  [1  ,  84]  =  0.684398
  [2  ,   2]  =  1.0
  [2  ,   7]  =  0.393775
  [2  ,  37]  =  0.0205692
  [2  ,  66]  =  0.537647
  [2  ,  95]  =  0.371229
  [3  ,   3]  =  1.0
  [3  ,   8]  =  0.165872
  [3  ,  40]  =  0.335283
  [3  ,  47]  =  0.256675
  [3  ,  58]  =  0.495783
  [3  ,  78]  =  0.959274
  [3  ,  84]  =  0.63355
  [3  ,  89]  =  0.0644833
  [3  ,  95]  =  0.355326
  [4  ,   4]  =  1.0
  [4  ,  27]  =  0.932093
  [4  ,  39]  =  0.826629
  [4  ,  46]  =  0.659616
  [4  ,  53]  =  0.730766
  ⋮
  [97 ,  14]  =  0.997345
  [97 ,  88]  =  0.814524
  [97 ,  97]  =  1.0
  [98 ,  13]  =  0.976985
  [98 ,  16]  =  0.865208
  [98 ,  24]  =  0.719739
  [98 ,  51]  =  0.592326
  [98 ,  61]  =  0.788088
  [98 ,  68]  =  0.496462
  [98 ,  77]  =  0.208691
  [98 ,  78]  =  0.469939
  [98 ,  87]  =  0.496266
  [98 ,  98]  =  1.0
  [99 ,  31]  =  0.888942
  [99 ,  51]  =  0.579006
  [99 ,  80]  =  0.372221
  [99 ,  82]  =  0.511931
  [99 ,  92]  =  0.191968
  [99 ,  96]  =  0.915728
  [99 ,  99]  =  1.0
  [100,  11]  =  0.771669
  [100,  15]  =  0.871662
  [100,  28]  =  0.402603
  [100,  92]  =  0.0595046
  [100, 100]  =  1.0

julia> rowPtr = CuVector{eltype(A_cpu_csr.rowptr), CUDA.UnifiedMemory}(A_cpu_csr.rowptr)
101-element CuArray{Int64, 1, CUDA.UnifiedMemory}:
   1
   7
  12
  21
  29
  34
  39
  46
  54
  60
  68
  77
  81
  89
  94
  98
 103
 110
 115
 124
 133
 138
 141
 147
 149
   ⋮
 459
 464
 468
 471
 476
 484
 490
 493
 498
 506
 515
 524
 531
 534
 537
 542
 549
 556
 560
 563
 571
 574
 584
 591
 596

julia> colVal = CuVector{eltype(A_cpu_csr.colval), CUDA.UnifiedMemory}(A_cpu_csr.colval)
595-element CuArray{Int64, 1, CUDA.UnifiedMemory}:
   1
  12
  19
  54
  79
  84
   2
   7
  37
  66
  95
   3
   8
  40
  47
  58
  78
  84
  89
  95
   4
  27
  39
  46
  53
   ⋮
  14
  88
  97
  13
  16
  24
  51
  61
  68
  77
  78
  87
  98
  31
  51
  80
  82
  92
  96
  99
  11
  15
  28
  92
 100

julia> nzVal  = CuVector{eltype(A_cpu_csr.nzval), CUDA.UnifiedMemory}(T.(A_cpu_csr.nzval))
595-element CuArray{Float64, 1, CUDA.UnifiedMemory}:
 1.0
 0.9419006148158842
 0.8359372756385358
 0.23207731655633967
 0.8212512726686211
 0.6843978774141242
 1.0
 0.39377516481792096
 0.020569239589359523
 0.5376474817496429
 0.3712292737100772
 1.0
 0.16587215790934517
 0.3352828993596362
 0.2566752647407271
 0.4957831809713593
 0.9592743664777489
 0.6335502177953432
 0.06448333949044249
 0.3553264016878862
 1.0
 0.9320934111824593
 0.8266287412165414
 0.6596158483158374
 0.7307662431644291
 ⋮
 0.9973450322725584
 0.8145235917870958
 1.0
 0.976985204969027
 0.8652082192479256
 0.7197390349689757
 0.5923256200255081
 0.7880884237319915
 0.4964615779473832
 0.20869080787992955
 0.4699389018381821
 0.49626647838376037
 1.0
 0.8889422805864055
 0.5790055975320247
 0.3722209529975524
 0.5119312004444105
 0.19196803033138043
 0.9157283595838851
 1.0
 0.7716693371121929
 0.871662421941884
 0.40260326157274373
 0.05950463737310885
 1.0

julia> A_gpu = CuSparseMatrixCSR(rowPtr, colVal, nzVal, (size(A_cpu_csr, 1), size(A_cpu_csr, 2)))
100×100 CuSparseMatrixCSR{Float64, Int32} with 595 stored entries:
⎡⠑⢄⠀⠢⠀⠈⠀⠀⠀⠁⠀⠀⠀⡀⠀⠀⠀⠀⠂⡠⠀⠀⢀⠄⠀⠀⡈⠀⠠⠀⠀⠀⠐⠀⠀⠀⡀⠀⠠⢁⠀⠨⡀⠀⠄⠀⠀⠆⠀⠀⎤
⎢⠀⠀⠑⣄⠀⠀⠀⢀⠀⠀⡀⠐⠀⠀⠀⠐⠀⠄⠄⠀⠀⠀⠐⠄⠀⠈⢀⠀⠀⠀⠈⠈⠀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠠⠀⣀⠀⠀⠁⠀⠦⎥
⎢⠠⠢⡀⠀⠑⢄⠒⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠁⠁⠀⠠⠀⠓⠀⠀⠀⠂⠀⠁⠠⠀⠀⠠⠀⠀⠀⠀⠀⠀⠄⠀⡀⠄⠀⠀⡀⠠⠀⠀⠑⎥
⎢⡀⠀⠂⠀⠀⠀⠑⢔⠐⠀⠈⠀⠀⠈⠀⠀⠀⠠⠀⠀⠠⠀⠀⠈⠀⢠⠀⠀⠀⡀⠀⠈⠀⠀⠀⠈⠀⠈⠀⠁⠀⠀⠀⠀⠀⡀⠂⠀⠀⠀⎥
⎢⠀⠀⠁⠀⠐⡀⠀⠠⠑⢄⠀⠀⠄⠀⠀⠀⠀⠀⠀⠀⠀⠌⠀⠀⠠⠀⠂⠐⠀⠀⠀⠀⠀⢀⢀⡨⡤⠀⠈⠂⠀⡀⠀⢀⢨⠀⠈⠀⠀⠀⎥
⎢⠠⠠⠀⠀⠀⠈⠀⡀⠀⠑⠑⢄⠀⠀⠠⠀⠀⠠⠀⠀⠀⠀⠀⠀⠀⠁⠀⠀⠀⠠⠀⠀⠀⠐⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠁⠀⠀⠀⠀⎥
⎢⠠⠁⠀⠀⠠⠀⠀⠀⠀⠀⠀⠁⠓⢄⠀⠀⠀⠀⠀⡀⠀⠂⠀⠀⠀⠄⠀⠠⠀⠀⠀⠀⡀⠠⠠⠀⠀⠠⠀⠈⠀⠐⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠈⠀⠀⠐⠀⠀⠈⠀⠀⠀⠀⠀⠑⢄⠀⠀⢀⠀⠁⠀⠀⠀⠠⠀⠄⡀⠀⠈⠀⠀⠄⠀⠀⠂⠀⠂⢀⡀⠠⠀⠈⢀⠀⢀⡀⠀⠀⢀⎥
⎢⠀⡀⠀⠀⢄⠀⠀⠀⠄⠀⠄⠀⠂⠀⠁⡐⠑⢄⠐⠀⠀⠁⠀⠀⠀⠀⠄⡀⠀⠀⠀⠁⠀⠂⠀⠁⠀⢀⠀⠀⠀⠀⡀⠀⢀⠀⠀⠀⠀⢀⎥
⎢⠀⢀⠄⣀⠀⡀⠠⠀⠀⠀⠀⠀⠀⢀⠠⠠⠀⠁⠑⢄⠀⠀⠀⠀⠀⠐⠂⠀⠀⠀⠀⠀⠀⠀⠀⠀⡀⠄⠀⠀⠀⠀⠀⠀⠀⠄⠀⠠⠀⠀⎥
⎢⠐⠢⠀⠄⠀⠀⠠⠀⠀⠀⠀⠁⠁⠂⠇⠀⠀⠀⠀⠀⢑⢌⠀⠀⠀⡈⠁⠀⠁⠀⠀⠄⡀⠂⠁⠀⠀⠀⠀⠀⢐⠀⠁⠀⠈⠆⠀⠀⠠⠀⎥
⎢⠁⠁⠐⠀⠂⠀⠀⡀⠀⠀⠀⠀⠢⠈⠐⠁⠀⠀⠀⠀⠀⠈⡑⢄⠀⠠⠀⠀⠀⠀⠀⠀⠀⠀⠁⠠⠀⠈⠕⠀⠀⠀⠀⠀⠀⡐⠀⠀⠀⠀⎥
⎢⠔⠀⡀⠀⠀⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠆⠠⠀⠀⠀⠀⠀⠀⡀⠑⢄⡀⠀⠀⠀⠂⠀⠐⢂⠀⠀⠀⢀⠀⠀⠀⢀⠑⠀⠀⠀⡀⠀⠀⠈⎥
⎢⠀⠀⠀⡀⠁⠀⠀⠀⠀⠀⢀⠀⠀⠈⠀⠀⠀⠀⠀⠀⢀⠀⠐⠀⠀⠀⠱⢄⠀⠀⠀⡀⠠⠈⠄⠀⠉⠐⠀⠀⠡⠊⠄⠈⠀⠀⠀⡁⠀⠀⎥
⎢⠀⡀⠀⠀⠐⠀⢀⠄⠀⠀⠀⠈⠀⠀⢀⠀⠂⠐⠀⡀⠀⠀⠁⠀⠀⢈⠠⠀⠑⣄⡀⠂⠅⠁⠁⢂⠀⠀⠀⠀⠀⠀⠐⠀⠀⠈⠀⠅⠁⠀⎥
⎢⠀⠀⠅⠀⠊⠀⠁⠀⠐⡀⠀⠀⠀⠀⠁⠈⢠⠀⠀⠠⠀⠐⠀⠀⠀⠀⠂⠀⠂⠐⠑⢅⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡈⠐⡀⠂⠀⠀⠀⎥
⎢⠀⠂⠀⠀⠀⠀⢀⠠⠀⠀⠀⠀⠀⠀⡀⠀⠀⠀⠀⠀⠠⠈⠈⠀⠊⠀⠀⠀⠀⠀⠀⠠⠑⢄⡀⠀⠀⠀⠀⠂⠀⠀⠀⢈⢀⠀⠀⠀⠄⠀⎥
⎢⠀⠂⠠⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠄⠀⠀⠀⠀⠀⡠⠀⠂⠀⠀⠀⢠⠁⠀⠀⠐⠂⠑⢆⠀⠀⠠⠀⠐⠀⠀⠀⠀⠠⠄⡀⠀⠀⎥
⎢⠠⠀⠀⠀⠀⠀⠈⠀⠀⠀⠀⠃⠐⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⠁⠄⠀⠀⠂⠀⠀⠀⠀⠀⠐⠀⠀⠕⢔⠀⡀⠀⠀⠀⠀⠀⠀⠀⠤⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⢀⠀⠄⠂⠀⢀⠀⢀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠀⠄⠀⠈⠀⠀⠀⠈⠀⠀⠀⠁⠑⢄⠀⠀⠂⠀⠀⠀⠀⠐⠀⠀⎥
⎢⡀⠀⠀⠀⠀⠀⠂⠀⠄⠈⠈⠀⠀⠀⠀⠂⠀⠀⠀⠀⠀⠀⠀⠀⠀⢘⠐⡀⠀⠀⠀⡁⠀⠀⠀⠁⠀⠄⠀⠀⠑⢄⠀⠐⠈⠀⠀⠀⠀⠁⎥
⎢⠀⠠⠀⠠⠀⠀⠀⠀⠀⠀⠀⠂⡀⠀⠀⠀⠂⢈⠀⠬⠀⠀⠑⠩⠀⠁⢀⠀⠀⠀⠰⡀⠀⠆⠀⠀⠀⠀⠐⡁⡀⠀⠑⢄⠂⠀⠀⠂⠄⠀⎥
⎢⠀⠆⠈⠀⠀⠀⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠀⠠⠀⠀⠀⠀⠀⠐⠀⠀⠀⠀⠀⠀⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡁⠀⠑⢄⢠⠄⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠄⠀⠀⠁⠈⠀⠀⠀⠀⢀⠄⡀⠀⠀⠀⠀⠀⠀⠀⢀⠐⠂⠈⠀⢁⠈⠈⠀⠀⢀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡑⢄⡂⠀⎥
⎣⠀⠀⠀⠀⠀⡀⠊⡐⠀⠀⠀⠐⠀⢀⠀⠄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠆⠀⠀⠀⠀⠂⠀⠀⠐⠀⠀⠀⠀⠒⠠⠠⠀⠀⠊⠀⢠⠀⠠⠑⢄⎦

julia> x_gpu = T.(cu(x_cpu; unified=true))
100-element CuArray{Float64, 1, CUDA.UnifiedMemory}:
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 ⋮
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0

julia> b_gpu = T.(cu(b_cpu; unified=true))
100-element CuArray{Float64, 1, CUDA.UnifiedMemory}:
 0.6885662078857422
 0.19811876118183136
 0.3463665843009949
 0.5088338851928711
 0.5786626935005188
 0.3527742028236389
 0.590097963809967
 0.6028531193733215
 0.5654492974281311
 0.975313127040863
 0.6991729736328125
 0.31716588139533997
 0.17102549970149994
 0.2767346203327179
 0.3544235825538635
 0.3042399287223816
 0.1599130481481552
 0.47176724672317505
 0.556181013584137
 0.9483237266540527
 0.5645405054092407
 0.5958035588264465
 0.07342114299535751
 0.7556286454200745
 0.4832138419151306
 ⋮
 0.9282295107841492
 0.9285739064216614
 0.5164573788642883
 0.7207266688346863
 0.09463237226009369
 0.5399411916732788
 0.9454120993614197
 0.12503908574581146
 0.7671582102775574
 0.2466679811477661
 0.9779777526855469
 0.9353416562080383
 0.824032187461853
 0.46944230794906616
 0.642370879650116
 0.9165357947349548
 0.6219422817230225
 0.23355454206466675
 0.20615164935588837
 0.40546366572380066
 0.8605594038963318
 0.6230349540710449
 0.27935782074928284
 0.4203110337257385
 0.09211910516023636

julia> solver = CudssSolver(A_gpu, "G", 'F')
CudssSolver{Float64, Int32}(CudssMatrix{Float64, Int32}(Float64, Int32, Ptr{CUDSS.cudssMatrix} @0x000000003d43e9e0, 1, 100, 100, 595), CudssConfig(Ptr{CUDSS.cudssConfig} @0x000000003c342f50), CudssData(Ptr{CUDSS.cudssContext} @0x000000003d25c790, Ptr{CUDSS.cudssData} @0x000000003f4307b0), true, PtrOrCuPtr{Nothing}(0x0000000000000000), Base.RefValue{Int32}(1), Base.RefValue{Int64}(1), Base.RefValue{Float64}(6.33035379135417e-310), Base.RefValue{Tuple{Int32, Int32}}((40, 0)), Base.RefValue{Tuple{Int64, Int64, Int64}}((128127724104032, 128127724099488, 128127724099680)), Base.RefValue{CUDSS.cudssAlgType_t}(CUDSS.CUDSS_ALG_1), Base.RefValue{CUDSS.cudssPivotType_t}(CUDSS.CUDSS_PIVOT_NONE), Base.RefValue{Ptr{CUDSS.cudssMatrix}}(Ptr{CUDSS.cudssMatrix} @0x0000000000000003), 0x0000000000000000, Base.RefValue{UInt64}(0x0000000000000001))

julia> cudss("analysis", solver, x_gpu, b_gpu)

julia> cudss("factorization", solver, x_gpu, b_gpu)
false

julia> cudss("solve", solver, x_gpu, b_gpu)

julia> x_gpu
100-element CuArray{Float64, 1, CUDA.UnifiedMemory}:
 -1.7632130781344415
 -0.24535435421854
 -0.511328665306535
 -1.1103600346022215
 -2.4018960066508237
  0.5678030294588474
  2.3090438464831045
 -2.6571946290597888
 -2.0716878705137423
  3.74779229599057
  1.683947176142427
  1.336091744677562
 -2.9681423141175443
  0.36207211979876086
  1.008322574205535
  1.4990297181671994
  0.6739180997108912
  1.6337736196857537
 -2.798819744636641
 -0.7597351519495357
  1.4004914033284868
  1.4519608324286235
  3.192487365934526
  0.2771279147244061
 -1.0683215746279329
  ⋮
  0.44184304933280316
  2.1388763858140734
  0.01182353557714444
  1.2047259359176836
 -0.48418996611998
 -0.9111688327514906
  0.6813551564456366
 -0.8009086215815431
  3.1625092732138103
 -2.0443971519749935
 -1.9170069218171
  0.5171882449218994
  1.8381057584213027
  0.3151084072854057
  1.1935982665716998
 -0.001480128376457489
  3.408901283418406
  0.438320267302093
  1.3223258073558013
  0.5732366743355825
 -2.3816694620729524
 -1.235256380368507
 -0.7215855911415417
  2.050318938453117
 -2.4654097607846635

julia> norm(b_gpu - A_gpu * x_gpu)
1.5257521659920597e-14

I don't have any issue here.
I use a H200.

[amontoison]

@albertomercurio I checked the details and we don't have any control on the buffer type in cuDSS, all storage is allocated internally.
We have a few options to use hybrid_memory_mode directly from cuDSS but it is independent of CUDA.jl:
https://exanauts.github.io/CUDSS.jl/dev/options/#Hybrid-memory-mode

[albertomercurio]

I don't have any issue here.
I use a H200

My example was just showing how I use shared memory here, but if you increase the size from n=100 to a way larger value, I noticed that somehow it is not taking advantage of the unified memory that much.

@albertomercurio I checked the details and we don't have any control on the buffer type in cuDSS, all storage is allocated internally.

Ok so CUDSS might allocate internally device memory?

[amontoison]

Yes, it seems to be by default: https://docs.nvidia.com/cuda/cudss/advanced_features.html#hybrid-host-device-memory-mode