我正在尝试使用PyCUDA来连接稀疏cuSOLVER例程cusolverSpDcsrlsvqr()(> = CUDA 7.0),并且遇到了一些困难:我试图以将密集cuSolver例程包裹在scikits-cuda(https:/中的方式)包装这些方法/github.com/lebedov/scikits.cuda/blob/master/scikits/cuda/cusolver.py)。
但是,在调用cusolverSpDcsrlsvqr()函数时,代码因分段错误而崩溃。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中的cusolverSpXcsrissymHost()中的#0 0x00007fffd9e3b71a(1
)来自/ usr / local / cudar_xr_x_rx_x0_x(0 * 0764)的0x00007fffd9e3b71a来自/ usr / local / cuda / lib64 / libcus_analys0(764) / usr
/ local /
lib / lib中的/local/cuda/lib64/libcusolver.so#3 0x00007fffd9f160a0在cusolverXcsrqr_analysis()中/usr/local/cuda/lib64/libcusolver.so#4 0x00007fffd9f28d78在cusolverSpScsrlsvqr()中。
这很奇怪,因为我既不调用cusolverSp S csrlsvqr(),也不认为它应该调用宿主函数(cusolverSpXcsrissym Host)。
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))
设置descrA
于ctypes.c_void_p()
和更换cusparseMatDescr_t
的cusolverSpDcsrlsvqr
包装与ctypes.c_void_p
应解决的问题。
本文收集自互联网,转载请注明来源。
如有侵权,请联系[email protected] 删除。
我来说两句