使用 scikit-cuda 通过 cuSOLVER 计算奇异值分解

Question

我正在尝试将 scikit-cuda 的包装器用于 cuSOLVER 函数，特别是我想执行 cusolverDnSgesvd 来计算实数矩阵上的全矩阵单精度 SVD。

使用代码here 和here 作为参考，我设法做到了这一点：

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

from skcuda import cusolver


handle = cusolver.cusolverDnCreate()

m = 50
n = 25

a = np.asarray(np.random.random((m, n)))
a_gpu = gpuarray.to_gpu(a)

ldu = m
ldvt = n

s_gpu = gpuarray.empty(min(m, n), np.float32)
u_gpu = gpuarray.empty((ldu, m), np.float32)
vh_gpu = gpuarray.empty((n, n), np.float32)

work_size = cusolver.cusolverDnSgesvd_bufferSize(handle, m, n)

work = gpuarray.empty((m,n), np.float32)

u_gpu, s_gpu, vh_gpu = cusolver.cusolverDnSgesvd(
    handle=handle,
    jobu='A',
    jobvt='A',
    m=m,
    n=n,
    A=a,
    lda=m,
    S=s_gpu,
    U=u_gpu,
    ldu=ldu,
    VT=vh_gpu,
    ldvt=ldvt,
    Work=work,
    Lwork=work_size,
    rwork=None,
    devInfo=0
)

但是代码不起作用，可能是因为我搞砸了类型。

Traceback (most recent call last):
  File "/home/vektor/PycharmProjects/yancut/test_svd.py", line 44, in <module>
    devInfo=0
  File "/home/vektor/anaconda3/lib/python3.4/site-packages/skcuda/cusolver.py", line 577, in cusolverDnSgesvd
    int(A), lda, int(S), int(U),
TypeError: only length-1 arrays can be converted to Python scalars

我应该如何提供所有参数以便 SVD 以正确的方式执行？

更新1： 在使用this 问题作为参考后，我编辑了我的代码，但出现了一个新错误。

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

import ctypes

from skcuda import cusolver

rows = 20
cols = 10

a = np.asarray(np.random.random((rows, cols)))
a_gpu = gpuarray.to_gpu(a.copy())
lda = rows

u_gpu = gpuarray.empty((rows, rows), np.float32)
v_gpu = gpuarray.empty((cols, cols), np.float32)
s_gpu = gpuarray.empty(cols, np.float32)
devInfo = gpuarray.zeros(1, np.int32)

handle = cusolver.cusolverDnCreate()

worksize = cusolver.cusolverDnSgesvd_bufferSize(handle, rows, cols)
print("SIZE", worksize)

Workspace = gpuarray.empty(worksize, np.float32)

svd_status = cusolver.cusolverDnSgesvd(
    handle=handle,
    jobu='A',
    jobvt='A',
    m=rows,
    n=cols,
    A=a_gpu.ptr,
    lda=rows,
    S=s_gpu.ptr,
    U=u_gpu.ptr,
    ldu=rows,
    VT=v_gpu.ptr,
    ldvt=cols,
    Work=Workspace.ptr,
    Lwork=worksize,
    rwork=Workspace.ptr,
    devInfo=devInfo.ptr
)
status = cusolver.cusolverDnDestroy(handle)

我遇到了一个新错误

Traceback (most recent call last):
  File "/home/vektor/PycharmProjects/yancut/test_svd.py", line 53, in <module>
    devInfo=devInfo.ptr
  File "/home/vektor/anaconda3/lib/python3.4/site-packages/skcuda/cusolver.py", line 579, in cusolverDnSgesvd
    Lwork, int(rwork), int(devInfo))
ctypes.ArgumentError: argument 2: <class 'TypeError'>: wrong type

现在看来我对 devInfo 做错了什么

Answer 1

从documentation 看来，每个矩阵（所以A、S、U、VT）都需要作为设备指针传递。所以对于 PyCUDA gpuarrays，传递 A.ptr 而不是 A。等等，它应该可以工作。