Getrs Function Of Cusolver Over Pycuda Doesn't Work Properly

I'm trying to make a pycuda wrapper inspired by scikits-cuda library for some operations provided in the new cuSolver library of Nvidia. I want to solve a linear system of the form

Solution 1:

Here is a full working example of how to use the library; the result is validated against that obtained using numpy's built-in solver:

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)