我写了一个Cython程序,调用Intel MKL进行矩阵乘法,目的是使它并行。它是基于一个旧的SO post链接到BLAS的,使用了一堆我从未见过的Cython方法,但它运行起来比NumPy(也链接到MKL)慢得多。为了加快速度,我使用了典型的Memoryview格式(它使用ndarray
np.float64_t
数据类型进行一些操作)。但现在它不再使用double[::1]
内存视图。以下是生成的错误:
'type cast': cannot convert from '__Pyx_memviewslice' to 'double *'
由于类型转换不起作用,MKL函数只能看到5个参数中的3个:
error C2660: 'cblas_ddot': function does not take 3 arguments
以下是.PYX代码:
import numpy as np
cimport numpy as np
cimport cython
from cython cimport view
from cython.parallel cimport prange #this is your OpenMP portion
from openmp cimport omp_get_max_threads #only used for getting the max # of threads on the machine
cdef extern from "mkl_cblas.h" nogil: #import a function from Intel's MKL library
double ddot "cblas_ddot"(int N,
double *X,
int incX,
double *Y,
int incY)
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.cdivision(True)
cpdef matmult(double[:,::1] A, double[:,::1] B):
cdef int Ashape0=A.shape[0], Ashape1=A.shape[1], Bshape0=B.shape[0], Bshape1=B.shape[1], Arowshape0=A[0,:].shape[0] #these are defined here as they aren't allowed in a prange loop
if Ashape1 != Bshape1:
raise TypeError('Inner dimensions are not consistent!')
cdef int i, j
cdef double[:,::1] out = np.zeros((Ashape0, Bshape1))
cdef double[::1] A_row = np.zeros(Ashape0)
cdef double[:] B_col = np.zeros(Bshape1) #no idea why this is not allowed to be [::1]
cdef int Arowstrides = A_row.strides[0] // sizeof(double)
cdef int Bcolstrides = B_col.strides[0] // sizeof(double)
cdef int maxthreads = omp_get_max_threads()
for i in prange(Ashape0, nogil=True, num_threads=maxthreads, schedule='static'): # to use all cores
A_row = A[i,:]
for j in range(Bshape1):
B_col = B[:,j]
out[i,j] = ddot(Arowshape0, #call the imported Intel MKL library
<double*>A_row,
Arowstrides,
<double*>B_col,
Bcolstrides)
return np.asarray(out)
我相信这是很容易被人指出的。如果你看到了可以改进的地方,请给出建议-这是被砍掉的,我认为甚至不需要I/j循环。关于:https://gist.github.com/JonathanRaiman/f2ce5331750da7b2d4e9的最干净的例子,我最后编译它的速度实际上快得多(2倍),但没有给出任何结果,所以我将把它放在另一篇文章中(这里:Calling BLAS / LAPACK directly using the SciPy interface and Cython - also how to add MKL)
非常感谢。在
要从memoryview获取指针,需要获取第一个元素的地址
相关问题 更多 >
编程相关推荐