0

我正在尝试为 Nvidia 的新 cuSolver 库中提供的一些操作制作一个受 scikits-cuda 库启发的 pycuda 包装器。我想通过 LU 分解求解 AX=B 形式的线性系统，首先使用 scikits-cuda 中的cublasSgetrfBatched方法执行该方法，这给了我分解 LU；然后通过该因式分解，我想使用我想要包装的cuSolve中的 cusolverDnSgetrs 来求解系统，当我执行计算返回状态 3 时，假设给我答案的矩阵不会改变，但 *devInfo 为零，查看 cusolver 的文档说：

CUSOLVER_STATUS_INVALID_VALUE=向函数传递了一个不支持的值或参数（例如，一个负向量大小）。

libcusolver.cusolverDnSgetrs.restype=int
libcusolver.cusolverDnSgetrs.argtypes=[_types.handle,
                                   ctypes.c_char,
                                   ctypes.c_int,
                                   ctypes.c_int,
                                   ctypes.c_void_p,
                                   ctypes.c_int,
                                   ctypes.c_void_p,
                                   ctypes.c_void_p,
                                   ctypes.c_int,
                                   ctypes.c_void_p]

"""
handle is the handle pointer given by calling cusolverDnCreate() from cuSolver
LU is the LU factoriced matrix given by cublasSgetrfBatched() from scikits
P is the pivots matrix given by cublasSgetrfBatched()
B is the right hand matix from AX=B
"""
def cusolverSolveLU(handle,LU,P,B):
   rows_LU ,cols_LU = LU.shape
   rows_B, cols_B = B.shape
   B_gpu = gpuarray.to_gpu(B.astype('float32'))
   info_gpu = gpuarray.zeros(1, np.int32)

   status=libcusolver.cusolverDnSgetrs(
               handle, 'n', rows_LU, cols_B,
               int(LU.gpudata), cols_LU,
               int(P.gpudata), int(B_gpu.gpudata),
               cols_B, int(info_gpu.gpudata))
   print info_gpu   
   print status

handle= cusolverCreate() #get the initialization of cusolver
LU, P = cublasLUFactorization(...)
B = np.asarray(np.random.rand(3, 3), np.float32)
cusolverSolveLU(handle,LU,P,B)

输出：

[0]

3

我做错了什么？

4

1 回答 1

3

这是如何使用该库的完整工作示例；使用 numpy 的内置求解器对结果进行验证：

import ctypes

import numpy as np
import pycuda.autoinit
import pycuda.gpuarray as gpuarray

CUSOLVER_STATUS_SUCCESS = 0

libcusolver = ctypes.cdll.LoadLibrary('libcusolver.so')

libcusolver.cusolverDnCreate.restype = int
libcusolver.cusolverDnCreate.argtypes = [ctypes.c_void_p]

def cusolverDnCreate():
    handle = ctypes.c_void_p()
    status = libcusolver.cusolverDnCreate(ctypes.byref(handle))
    if status != CUSOLVER_STATUS_SUCCESS:
        raise RuntimeError('error!')
    return handle.value

libcusolver.cusolverDnDestroy.restype = int
libcusolver.cusolverDnDestroy.argtypes = [ctypes.c_void_p]

def cusolverDnDestroy(handle):
    status = libcusolver.cusolverDnDestroy(handle)
    if status != CUSOLVER_STATUS_SUCCESS:
        raise RuntimeError('error!')

libcusolver.cusolverDnSgetrf_bufferSize.restype = int
libcusolver.cusolverDnSgetrf_bufferSize.argtypes = [ctypes.c_void_p,
                                                    ctypes.c_int,
                                                    ctypes.c_int,
                                                    ctypes.c_void_p,
                                                    ctypes.c_int,
                                                    ctypes.c_void_p]

def cusolverDnSgetrf_bufferSize(handle, m, n, A, lda, Lwork):
    status = libcusolver.cusolverDnSgetrf_bufferSize(handle, m, n,
                                                     int(A.gpudata),
                                                     n, ctypes.pointer(Lwork))
    if status != CUSOLVER_STATUS_SUCCESS:
        raise RuntimeError('error!')

libcusolver.cusolverDnSgetrf.restype = int
libcusolver.cusolverDnSgetrf.argtypes = [ctypes.c_void_p,
                                         ctypes.c_int,
                                         ctypes.c_int,
                                         ctypes.c_void_p,
                                         ctypes.c_int,
                                         ctypes.c_void_p,
                                         ctypes.c_void_p,
                                         ctypes.c_void_p]

def cusolverDnSgetrf(handle, m, n, A, lda, Workspace, devIpiv, devInfo):
    status = libcusolver.cusolverDnSgetrf(handle, m, n, int(A.gpudata),
                                          lda,
                                          int(Workspace.gpudata),
                                          int(devIpiv.gpudata),
                                          int(devInfo.gpudata))
    if status != CUSOLVER_STATUS_SUCCESS:
        raise RuntimeError('error!')

libcusolver.cusolverDnSgetrs.restype = int
libcusolver.cusolverDnSgetrs.argtypes = [ctypes.c_void_p,
                                         ctypes.c_int,
                                         ctypes.c_int,
                                         ctypes.c_int,
                                         ctypes.c_void_p,
                                         ctypes.c_int,
                                         ctypes.c_void_p,
                                         ctypes.c_void_p,
                                         ctypes.c_int,
                                         ctypes.c_void_p]

def cusolverDnSgetrs(handle, trans, n, nrhs, A, lda, devIpiv, B, ldb, devInfo):
    status = libcusolver.cusolverDnSgetrs(handle, trans, n, nrhs,
                                          int(A.gpudata), lda,
                                          int(devIpiv.gpudata), int(B.gpudata),
                                          ldb, int(devInfo.gpudata))
    if status != CUSOLVER_STATUS_SUCCESS:
        raise RuntimeError('error!')

if __name__ == '__main__':
    m = 3
    n = 3
    a = np.asarray(np.random.rand(m, n), np.float32)
    a_gpu = gpuarray.to_gpu(a.T.copy())
    lda = m
    b = np.asarray(np.random.rand(m, n), np.float32)
    b_gpu = gpuarray.to_gpu(b.T.copy())
    ldb = m

    handle = cusolverDnCreate()
    Lwork = ctypes.c_int()

    cusolverDnSgetrf_bufferSize(handle, m, n, a_gpu, lda, Lwork)
    Workspace = gpuarray.empty(Lwork.value, dtype=np.float32)
    devIpiv = gpuarray.zeros(min(m, n), dtype=np.int32)
    devInfo = gpuarray.zeros(1, dtype=np.int32)

    cusolverDnSgetrf(handle, m, n, a_gpu, lda, Workspace, devIpiv, devInfo)
    if devInfo.get()[0] != 0:
        raise RuntimeError('error!')
    CUBLAS_OP_N = 0
    nrhs = n
    devInfo = gpuarray.zeros(1, dtype=np.int32)
    cusolverDnSgetrs(handle, CUBLAS_OP_N, n, nrhs, a_gpu, lda, devIpiv, b_gpu, ldb, devInfo)

    x_cusolver = b_gpu.get().T
    cusolverDnDestroy(handle)

    x_numpy = np.linalg.solve(a, b)
    print np.allclose(x_numpy, x_cusolver)
于 2015-04-22T16:31:40.633 回答