Arrayfire C++ 稀疏矩阵乘法导致访问冲突

Arrayfire C++ sparse matrix multiplication causes access violation

提问人：Daniel Rohrbach 提问时间：7/3/2023 最后编辑：Robert CrovellaDaniel Rohrbach 更新时间：7/11/2023 访问量：111

问：

我正在尝试在 arrayfire 中使用稀疏矩阵，但我在 array fire dll 中的某个地方遇到了访问冲突。我用了几个例子，但总是得到相同的结果。我可以在 CUDA 和 CPU 后端看到错误。

法典：

这是我尝试运行的代码

af::info();
    af_print(af::randu(5, 4));

    float v[]   = {5, 8, 3, 6};
    int r[]     = {0, 0, 2, 3, 4};
    int c[]     = {0, 1, 2, 1};
    const int M = 4, N = 4, nnz = 4;
    af::array vals    = af::array(af::dim4(nnz), v);
    af::array row_ptr = af::array(af::dim4(M + 1), r);
    af::array col_idx = af::array(af::dim4(nnz), c);
 
    af_print(vals);
    af_print(row_ptr);
    af_print(col_idx);

    // Create sparse array (CSR) from af::arrays containing values,
    // row pointers, and column indices.
    auto sparseM = af::sparse(M, N, vals, row_ptr, col_idx, AF_STORAGE_CSR);
    af_print(sparseM);

    auto res = sparseM * sparseM;
    af_print(res);

我的配置：

在 Windows 11 上运行。在 VSCode 中使用 cmake 使用 Visual Studio Build 工具 2022 进行编译。（有关系统的详细信息可以在下面的 ArrayFire 输出中找到）。我的系统有两个 GPU，NVIDIA 和一个 Intel

[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DependencyModule.cpp(104) ] Found: forge.dll
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(217) ] Found 5 OpenCL platforms
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(229) ] Found 1 devices on platform NVIDIA CUDA
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(234) ] Found device NVIDIA GeForce GTX 1050 Ti with Max-Q Design on platform NVIDIA CUDA
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(229) ] Found 1 devices on platform Intel(R) OpenCL HD Graphics
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(234) ] Found device Intel(R) UHD Graphics 630 on platform Intel(R) OpenCL HD Graphics
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(229) ] Found 1 devices on platform Intel(R) OpenCL
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(234) ] Found device Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz on platform Intel(R) OpenCL
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(229) ] Found 1 devices on platform Intel(R) FPGA Emulation Platform for OpenCL(TM)
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(234) ] Found device Intel(R) FPGA Emulation Device on platform Intel(R) FPGA Emulation Platform for OpenCL(TM)
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(229) ] Found 0 devices on platform Intel(R) FPGA SDK for OpenCL(TM)
[platform][1688317712][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\opencl\device_manager.cpp(239) ] Found 4 OpenCL devices

它将在倒数第二行失败。当与另一个非稀疏矩阵相乘时，这也将失败。auto res = sparseM * sparseM;

这是上面的输出：

[unified][1688317011][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(146) ] Found: afcuda.dll
Loaded 'C:\Windows\System32\DriverStore\FileRepository\nvdmi.inf_amd64_893ed8ff453738db\nvcuda64.dll'. 
[unified][1688317012][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(153) ] Device Count: 1.
[unified][1688317012][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(208) ] AF_DEFAULT_BACKEND: cuda
[platform][1688317031][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DependencyModule.cpp(101) ] Attempting to load: forge.dll
[platform][1688317031][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DependencyModule.cpp(104) ] Found: forge.dll
[mem][1688317031][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DefaultMemoryManager.cpp(128) ] memory[0].max_bytes: 14.7 GB
ArrayFire v3.8.3 (CPU, 64-bit Windows, build 987d5675a)
[0] Intel: Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz[mem][1688317044][10220] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cpu\memory.cpp(147) ] nativeAlloc:    1 KB 0x2182dabc380

         2 
         3 
         4 

col_idx
[4 1 1 1]
         0 
         1 
         2 
         1 

sparseM
Storage Format : AF_STORAGE_CSR
[4 4 1 1]
sparseM: Values
[4 1 1 1]
    5.0000 
    8.0000 
    3.0000 
    6.0000 

sparseM: RowIdx
[5 1 1 1]
2
         0 
         2 
         3 
         4 

sparseM: ColIdx
[4 1 1 1]
         0 
         1 
         2 
         1 

2
Exception thrown at 0x00007FFEF118ADCC (af.dll) in EigenSim.exe: 0xC0000005: Access violation reading location 0xFFFFFFFFFFFFFFFF.

在查看调试堆栈跟踪时，我得到：

af.dll!af_get_last_error(char * * str, __int64 * len) Line 46 (f:\buildbot\worker\win10-cuda-installer\build\src\api\unified\error.cpp:46)
af.dll!af::operator*(const af::array & lhs, const af::array & rhs) Line 939 (f:\buildbot\worker\win10-cuda-installer\build\src\api\cpp\array.cpp:939)
EigenSim.exe!testBackend() Line 27 (c:\Development\Eigensim\src\main\RunTestEigenSim.cpp:27)
EigenSim.exe!main() Line 37 (c:\Development\Eigensim\src\main\RunTestEigenSim.cpp:37)
EigenSim.exe!invoke_main() Line 79

这是使用 CUDA 后端时的输出

[unified][1688317714][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(146) ] Found: afcuda.dll
Loaded 'C:\Windows\System32\DriverStore\FileRepository\nvdmi.inf_amd64_893ed8ff453738db\nvcuda64.dll'. 
[unified][1688317714][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(153) ] Device Count: 1.
[unified][1688317714][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\api\unified\symbol_manager.cpp(208) ] AF_DEFAULT_BACKEND: cuda
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DependencyModule.cpp(101) ] Attempting to load: forge.dll
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DependencyModule.cpp(104) ] Found: forge.dll
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(494) ] CUDA Driver supports up to CUDA 12.2.0 ArrayFire CUDA Runtime 12.0.0
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(479) ] CUDA driver version(12.2.0) not part of the CudaToDriverVersion array. Please create an issue or a pull request on the ArrayFire repository to update the CudaToDriverVersion variable with this version of the CUDA runtime.

[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(562) ] Found 1 CUDA devices
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(590) ] Found device: NVIDIA GeForce GTX 1050 Ti with Max-Q Design (sm_61) (4 GB | ~2076.416015625 GFLOPs | 6 SMs)
[platform][1688317721][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(625) ] AF_CUDA_DEFAULT_DEVICE: 
[platform][1688317722][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\cuda\device_manager.cpp(644) ] Default device: 0(NVIDIA GeForce GTX 1050 Ti with Max-Q Design)
ArrayFire v3.8.3 (CUDA, 64-bit Windows, build 987d5675a)
Platform: CUDA Runtime 12.0, Driver: 12020
[0] NVIDIA GeForce GTX 1050 Ti with Max-Q Design, 4096 MB, CUDA Compute 6.1
[mem][1688317723][16000] [ F:\buildbot\worker\win10-cuda-installer\build\src\backend\common\DefaultMemoryManager.cpp(128) ] memory[0].max_bytes: 3 GB

我尝试过的事情

我尝试了所有后端 CUDA、GPU 和 OpenCL，但没有得到任何好的结果。对稀疏矩阵使用不同的存储格式将产生略微不同的效果。有时在创建稀疏矩阵时会失败，有时在对稀疏矩阵执行操作时会失败。

C++ CPU 稀疏矩阵 ArrayFire

评论

答：

3赞 Edwin Solis 7/11/2023 #1

从你的措辞来看，我假设你的意图是执行矩阵-矩阵乘法。

如果是这种情况，则应使用函数完成操作的代码。对于稀疏矩阵，该操作仅适用于稀疏矩阵作为第一个参数，而将密集矩阵作为第二个参数。您可以在 ArrayFire 文档中了解更多信息。这意味着您尝试实现的目标目前是不可能的。你必须使第二个成为密集矩阵;像这样的东西：af::matmul

auto res = af::matmul(sparseM, af::dense(sparseM));

您的代码出现错误，因为它执行的是按元素乘法，而不是矩阵乘法，并且稀疏矩阵不支持按元素乘法操作。

评论

0赞 Daniel Rohrbach 7/12/2023
非常感谢您回复我。你说的有道理。事实上，你是对的，我想要一个带有密集向量的 matmul，我让它真正起作用了。我认为更大的问题实际上是，我没有收到来自 arrayfire 的错误消息，或者后端它只是因访问违规而崩溃。我能够使用其他显然不允许的条件重现它，但会导致 arrayfire 崩溃并在我的配置上出现访问冲突。
1赞 Edwin Solis 7/12/2023
为避免崩溃，应正确处理异常。当 ArrayFire 出现错误时，它将引发异常。使用块来捕获错误并避免崩溃。c++ try {} catch(const af::exception&) {}
2赞 arrayfire 7/20/2023
GitHub 中正在跟踪同样的问题，此处为：github.com/arrayfire/arrayfire/issues/3460

上一个：Matlab mldivide 和 arrayfire 求解

下一个：如何使用 ArrayFire 批量 2D 卷积