JuliaGPU/CUDA.jl

[CUSPARSE] error in conversion of CuSparseMatrix{Int} to dense matrices

Open

#1,664 创建于 2022年11月6日

在 GitHub 查看
 (1 评论) (0 反应) (0 负责人)Julia (274 fork)batch import
bugcuda librariesgood first issue

仓库指标

Star
 (1,408 star)
PR 合并指标
 (平均合并 5天 5小时) (30 天内合并 16 个 PR)

描述

Cusparse matrices with float eltype are correctly converted to dense arrays:

julia> using CUDA, SparseArrays

julia> x = sparse([1, 2], [1, 2], [10.0, 20.0]) |> cu
2×2 CuSparseMatrixCSC{Float32, Int32} with 2 stored entries:
 10.0    ⋅ 
   ⋅   20.0

julia> CuArray(x)
2×2 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:
 10.0   0.0
  0.0  20.0

julia> Array(x)
2×2 Matrix{Float32}:
 10.0   0.0
  0.0  20.0

With Int eltype instead, we have the following StackOverflow errors:

julia> x = sparse([1,2], [1,2], [10, 20]) |> cu
2×2 CuSparseMatrixCSC{Int64, Int32} with 2 stored entries:
 10   ⋅
  ⋅  20

julia> CuArray(x)
ERROR: StackOverflowError:
Stacktrace:
     [1] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/libcuda.jl:3136 [inlined]
     [2] cuMemAllocAsync(dptr::Base.RefValue{CuPtr{Nothing}}, bytesize::Int64, hStream::CuStream)
       @ CUDA ~/.julia/dev/CUDA/lib/utils/call.jl:26
     [3] #alloc#1
       @ ~/.julia/dev/CUDA/lib/cudadrv/memory.jl:83 [inlined]
     [4] actual_alloc(bytes::Int64; async::Bool, stream::CuStream)
       @ CUDA ~/.julia/dev/CUDA/src/pool.jl:39
     [5] macro expansion
       @ ~/.julia/dev/CUDA/src/pool.jl:222 [inlined]
     [6] macro expansion
       @ ./timing.jl:382 [inlined]
     [7] #_alloc#164
       @ ~/.julia/dev/CUDA/src/pool.jl:303 [inlined]
     [8] #alloc#163
       @ ~/.julia/dev/CUDA/src/pool.jl:289 [inlined]
     [9] alloc
       @ ~/.julia/dev/CUDA/src/pool.jl:283 [inlined]
    [10] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(#unused#::UndefInitializer, dims::Tuple{Int64, Int64})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:85
    [11] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:334
    [12] (CuArray{Int64, 2})(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:339
    [13] copyto!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA.CUSPARSE ~/.julia/dev/CUDA/lib/cusparse/conversions.jl:596
    [14] copyto_axcheck!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./abstractarray.jl:1127
    [15] Matrix{Int64}(x::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:626
    [16] (Array{Int64})(A::CuSparseMatrixCSC{Int64, Int32})
       @ Core ./boot.jl:484
    [17] convert(#unused#::Type{Array{Int64}}, a::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:617
    [18] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:335
--- the last 7 lines are repeated 11423 more times ---
 [79980] (CuArray{Int64, 2})(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:339
 [79981] copyto!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA.CUSPARSE ~/.julia/dev/CUDA/lib/cusparse/conversions.jl:596
 [79982] copyto_axcheck!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./abstractarray.jl:1127
 [79983] Matrix{Int64}(x::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:626

julia> Array(x)
ERROR: StackOverflowError:
Stacktrace:
     [1] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/libcuda.jl:3136 [inlined]
     [2] cuMemAllocAsync(dptr::Base.RefValue{CuPtr{Nothing}}, bytesize::Int64, hStream::CuStream)
       @ CUDA ~/.julia/dev/CUDA/lib/utils/call.jl:26
     [3] #alloc#1
       @ ~/.julia/dev/CUDA/lib/cudadrv/memory.jl:83 [inlined]
     [4] actual_alloc(bytes::Int64; async::Bool, stream::CuStream)
       @ CUDA ~/.julia/dev/CUDA/src/pool.jl:39
     [5] macro expansion
       @ ~/.julia/dev/CUDA/src/pool.jl:222 [inlined]
     [6] macro expansion
       @ ./timing.jl:382 [inlined]
     [7] #_alloc#164
       @ ~/.julia/dev/CUDA/src/pool.jl:303 [inlined]
     [8] #alloc#163
       @ ~/.julia/dev/CUDA/src/pool.jl:289 [inlined]
     [9] alloc
       @ ~/.julia/dev/CUDA/src/pool.jl:283 [inlined]
    [10] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(#unused#::UndefInitializer, dims::Tuple{Int64, Int64})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:85
    [11] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:334
    [12] (CuArray{Int64, 2})(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:339
    [13] copyto!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA.CUSPARSE ~/.julia/dev/CUDA/lib/cusparse/conversions.jl:596
    [14] copyto_axcheck!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./abstractarray.jl:1127
    [15] Matrix{Int64}(x::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:626
    [16] (Array{Int64})(A::CuSparseMatrixCSC{Int64, Int32})
       @ Core ./boot.jl:484
    [17] convert(#unused#::Type{Array{Int64}}, a::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:617
    [18] CuArray{Int64, 2, CUDA.Mem.DeviceBuffer}(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:335
--- the last 7 lines are repeated 11423 more times ---
 [79980] (CuArray{Int64, 2})(xs::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA ~/.julia/dev/CUDA/src/array.jl:339
 [79981] copyto!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ CUDA.CUSPARSE ~/.julia/dev/CUDA/lib/cusparse/conversions.jl:596
 [79982] copyto_axcheck!(dest::Matrix{Int64}, src::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./abstractarray.jl:1127
 [79983] Matrix{Int64}(x::CuSparseMatrixCSC{Int64, Int32})
       @ Base ./array.jl:626

Same errors occur with SparseMatrixCSR and with both Int32 and Int64.

Tested on CUDA.jl master and the following versioninfo

julia> CUDA.versioninfo()
CUDA runtime 11.8, artifact installation
CUDA driver 11.6
NVIDIA driver 510.39.1

Libraries: 
- CUBLAS: 11.11.3
- CURAND: 10.3.0
- CUFFT: 10.9.0
- CUSOLVER: 11.4.1
- CUSPARSE: 11.7.5
- CUPTI: 18.0.0
- NVML: 11.0.0+510.39.1

Toolchain:
- Julia: 1.8.2
- LLVM: 13.0.1
- PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4, 6.5, 7.0, 7.1, 7.2
- Device capability support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75, sm_80, sm_86

4 devices:
  0: Tesla V100-SXM2-32GB (sm_70, 2.550 GiB / 32.000 GiB available)
  1: Tesla V100-SXM2-32GB (sm_70, 21.359 GiB / 32.000 GiB available)
  2: Tesla V100-SXM2-32GB (sm_70, 28.537 GiB / 32.000 GiB available)
  3: Tesla V100-SXM2-32GB (sm_70, 9.699 GiB / 32.000 GiB available)

贡献者指南

研究方向
检查 CuSparseMatrixCSC 的 copyto! 函数在整数类型中的无限递归问题。
技术栈
julia
领域
backend
议题类型
缺陷
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 天
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
活跃
清晰度议题是否清楚说明期望改动、验收标准和下一步。
清晰
前置要求
JuliaCUDA
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
60