Question

我正在尝试使用PyCUDA连接稀疏的cuSOLVER例程 cusolverSpDcsrlsvqr（）（＆gt; = CUDA 7.0）并面临一些困难：我尝试使用与scikits-cuda（https://github.com/lebedov/scikits.cuda/blob/master/scikits/cuda/cusolver.py）中包含的密集cuSolver例程相同的方式包装方法。

但是，在调用cusolverSpDcsrlsvqr（）函数时，代码会因分段错误而崩溃。使用cuda-gdb（cuda-gdb --args python -m pycuda.debug test.py; run;bt）进行调试会产生以下堆栈跟踪，

来自/usr/local/cuda/lib64/libcusolver.so的cusolverSpXcsrissymHost（）中的＃0> 0x00007fffd9e3b71a 来自/usr/local/cuda/lib64/libcusolver.so的hsolverXcsrqr_zeroPivot（）中的＃1 0x00007fffd9df5237 来自/usr/local/cuda/lib64/libcusolver.so的hsolverXcsrqr_analysis_coletree（）中的＃2 0x00007fffd9e0c764 来自/usr/local/cuda/lib64/libcusolver.so的cusolverXcsrqr_analysis（）中的＃3 0x00007fffd9f160a0 来自/usr/local/cuda/lib64/libcusolver.so的cusolverSpScsrlsvqr（）中的＃4 0x00007fffd9f28d78

这很奇怪，因为我没有调用cusolverSp S csrlsvqr（），我认为它也不应该调用主机函数（cusolverSpXcsrissym 主机）。

这是我正在谈论的代码 - 感谢您的帮助：

# ### Interface cuSOLVER PyCUDA


import pycuda.gpuarray as gpuarray
import pycuda.driver as cuda
import pycuda.autoinit
import numpy as np
import scipy.sparse as sp
import ctypes


# #### wrap the cuSOLVER cusolverSpDcsrlsvqr() using ctypes

# cuSparse
_libcusparse = ctypes.cdll.LoadLibrary('libcusparse.so')

class cusparseMatDescr_t(ctypes.Structure):
    _fields_ = [
        ('MatrixType', ctypes.c_int),
        ('FillMode', ctypes.c_int),
        ('DiagType', ctypes.c_int),
        ('IndexBase', ctypes.c_int)
        ]
_libcusparse.cusparseCreate.restype = int
_libcusparse.cusparseCreate.argtypes = [ctypes.c_void_p]

_libcusparse.cusparseDestroy.restype = int
_libcusparse.cusparseDestroy.argtypes = [ctypes.c_void_p]

_libcusparse.cusparseCreateMatDescr.restype = int
_libcusparse.cusparseCreateMatDescr.argtypes = [ctypes.c_void_p]


# cuSOLVER
_libcusolver = ctypes.cdll.LoadLibrary('libcusolver.so')



_libcusolver.cusolverSpCreate.restype = int
_libcusolver.cusolverSpCreate.argtypes = [ctypes.c_void_p]

_libcusolver.cusolverSpDestroy.restype = int
_libcusolver.cusolverSpDestroy.argtypes = [ctypes.c_void_p]



_libcusolver.cusolverSpDcsrlsvqr.restype = int
_libcusolver.cusolverSpDcsrlsvqr.argtypes= [ctypes.c_void_p,
                                            ctypes.c_int,
                                            ctypes.c_int,
                                            cusparseMatDescr_t,
                                            ctypes.c_void_p,
                                            ctypes.c_void_p,
                                            ctypes.c_void_p,
                                            ctypes.c_void_p,
                                            ctypes.c_double,
                                            ctypes.c_int,
                                            ctypes.c_void_p,
                                            ctypes.c_void_p]


#### Prepare the matrix and parameters, copy to Device via gpuarray

# coo to csr
val = np.arange(1,5,dtype=np.float64)
col = np.arange(0,4,dtype=np.int32)
row = np.arange(0,4,dtype=np.int32)
A = sp.coo_matrix((val,(row,col))).todense()
Acsr = sp.csr_matrix(A)
b = np.ones(4)
x = np.empty(4)
print('A:' + str(A))
print('b: ' + str(b))


dcsrVal = gpuarray.to_gpu(Acsr.data)
dcsrColInd = gpuarray.to_gpu(Acsr.indices)
dcsrIndPtr = gpuarray.to_gpu(Acsr.indptr)
dx = gpuarray.to_gpu(x)
db = gpuarray.to_gpu(b)
m = ctypes.c_int(4)
nnz = ctypes.c_int(4)
descrA = cusparseMatDescr_t()
reorder = ctypes.c_int(0)
tol = ctypes.c_double(1e-10)
singularity = ctypes.c_int(99)


#create cusparse handle
_cusp_handle = ctypes.c_void_p()
status = _libcusparse.cusparseCreate(ctypes.byref(_cusp_handle))
print('status: ' + str(status))
cusp_handle = _cusp_handle.value

#create MatDescriptor
status = _libcusparse.cusparseCreateMatDescr(ctypes.byref(descrA))
print('status: ' + str(status))

#create cusolver handle
_cuso_handle = ctypes.c_void_p()
status = _libcusolver.cusolverSpCreate(ctypes.byref(_cuso_handle))
print('status: ' + str(status))
cuso_handle = _cuso_handle.value



print('cusp handle: ' + str(cusp_handle))
print('cuso handle: ' + str(cuso_handle))


### Call solver
_libcusolver.cusolverSpDcsrlsvqr(cuso_handle,
                                 m,
                                 nnz,
                                 descrA,
                                 int(dcsrVal.gpudata),
                                 int(dcsrIndPtr.gpudata),
                                 int(dcsrColInd.gpudata),
                                 int(db.gpudata),
                                 tol,
                                 reorder,
                                 int(dx.gpudata),
                                 ctypes.byref(singularity))

# destroy handles
status = _libcusolver.cusolverSpDestroy(cuso_handle)
print('status: ' + str(status))
status = _libcusparse.cusparseDestroy(cusp_handle)
print('status: ' + str(status))

Answer 1

将descrA设置为ctypes.c_void_p()并将cusparseMatDescr_t包装中的cusolverSpDcsrlsvqr替换为ctypes.c_void_p可以解决问题。

使用PyCUDA连接cuSOLVER-sparse

1 个答案: