cython memoryview not faster than ndarray - numpy

Question

I have a function written in regular numpy ndarray and another one with a typed memoryview. However, I couldn't get the memoryview version to work faster than the regular version (unlike many of the blogs such as memoryview benchmarks).
Any pointers / suggestions to increase the speed of the memoryview code versus the numpy alternative will be greatly appreciated! ... OR ... if anyone can point out any glaring reason why the memoryview version is not much faster than the regular numpy version
In the code below there are two functions, both of which takes in two vectors bi and xi and returns a matrix. The first function shrink_correl is the regular numpy version and the second function shrink_correl2 is the memoryview alternative (let the file be sh_cor.pyx).
# cython: boundscheck=False
# cython: wraparound=False
# cython: cdivision=True
cimport cython
cimport numpy as np
import numpy as np
from numpy cimport ndarray as ar
# -- ***this is the Regular Cython version*** -
cpdef ar[double, ndim=2, mode='c'] shrink_correl(ar[double, ndim=1, mode='c'] bi, ar[double, ndim=1, mode='c'] xi):
cdef:
int n_ = xi.shape[0]
int n__ = int(n_*(n_-1)/2)
ar[double, ndim=2, mode='c'] f = np.zeros([n__, n_+1])
int x__ = 0
ar[double, ndim=2, mode='c'] f1 = np.zeros([n_, n_+1])
ar[double, ndim=2, mode='c'] f2 = np.zeros([n__, n_+1])
ar[double, ndim=1, mode='c'] g = np.zeros(n_+1)
ar[double, ndim=1, mode='c'] s = np.zeros(n__)
ar[double, ndim=2, mode='c'] cori_ = np.zeros([n_, n_])
Py_ssize_t j, k
with nogil:
for j in range(0, n_-1):
for k in range(j+1, n_):
x__ += 1
f[x__-1, j] = bi[k]*xi[k]*1000
f[x__-1, k] = bi[j]*xi[j]*1000
f1 = np.dot(np.transpose(f), f)
with nogil:
for j in range(0, n_):
f1[n_, j] = xi[j]*1000
f1[j, n_] = f1[n_, j]
f2 = np.dot(f, np.linalg.inv(f1))
with nogil:
for j in range(0, n_):
g[j] = -bi[j]*xi[j]*1000
s = np.dot(f2, g)
with nogil:
for j in range(0, n_):
cori_[j, j] = 1.0
x__ = 0
with nogil:
for j in range(0, n_-1):
for k in range(j+1, n_):
x__ += 1
cori_[j, k] = s[x__-1]
cori_[k, j] = cori_[j, k]
return cori_
# -- ***this is the MemoryView Cython version*** -
cpdef ar[double, ndim=2, mode='c'] shrink_correl2(double[:] bi, double[:] xi):
cdef:
int n_ = xi.shape[0]
int n__ = int(n_*(n_-1)/2)
double[:, ::1] f = np.zeros([n__, n_+1])
int x__ = 0
double[:, ::1] f1 = np.zeros([n_, n_+1])
double[:, ::1] f2 = np.zeros([n__, n_+1])
double[:] g = np.zeros(n_+1)
double[:] s = np.zeros(n__)
double[:, ::1] cori_ = np.zeros([n_, n_])
ar[double, ndim=2, mode='c'] cori__ = np.zeros([n_, n_])
Py_ssize_t j, k
with nogil:
for j in range(0, n_-1):
for k in range(j+1, n_):
x__ += 1
f[x__-1, j] = bi[k]*xi[k]*1000
f[x__-1, k] = bi[j]*xi[j]*1000
f1 = np.dot(np.transpose(f), f)
with nogil:
for j in range(0, n_):
f1[n_, j] = xi[j]*1000
f1[j, n_] = f1[n_, j]
f2 = np.dot(f, np.linalg.inv(f1))
with nogil:
for j in range(0, n_):
g[j] = -bi[j]*xi[j]*1000
s = np.dot(f2, g)
with nogil:
for j in range(0, n_):
cori_[j, j] = 1.0
x__ = 0
with nogil:
for j in range(0, n_-1):
for k in range(j+1, n_):
x__ += 1
cori_[j, k] = s[x__-1]
cori_[k, j] = cori_[j, k]
cori__[:, :] = cori_
return cori__
This is compiled using the following setup.py code
from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext
import numpy as np
import os
ext_modules = [Extension('sh_cor', ['sh_cor.pyx'], include_dirs=[np.get_include(),
os.path.join(np.get_include(), 'numpy')],
define_macros=[('NPY_NO_DEPRECATED_API', None)],
extra_compile_args=['-O3', '-march=native', '-ffast-math', '-flto'],
libraries=['m']
)]
setup(
name="Sh Cor",
cmdclass={'build_ext': build_ext},
ext_modules=ext_modules
)
The code used to test the speeds is
import numpy as np
import sh_cor # this the library created by the setup.py file
import time
b = np.random.random(400)
b = b/np.sum(b)
x = np.random.random(400)-0.5
n = 10
t0 = time.time()
for i in range(n):
v1 = sh_cor.shrink_correl(b, x)
t1 = time.time()
print((t1-t0)/n)
t0 = time.time()
for i in range(n):
v2 = sh_cor.shrink_correl2(b, x)
t1 = time.time()
print((t1-t0)/n)
The output on my PC is:
0.7070999860763549 # regular numpy
0.6726999998092651 # memoryview
using memoryview (in the codes above) only gives me a 5% speed boost (unlike the huge speed boost in the blogs).

Answer 1

#uday Give me about a week as I'm computer less, but here's where to speed things up to get you started: 1) instead of attaching the gil using np.transpose create a memoryview identical to what you want to transpose BEFORE any loops (I.e. you'll have variable f declared as a memoryview that won't need the gil and just create a view on that f_t, I.e. cdef double[:, ::1] f_T = np.transpose(f) or just =f.T.
2) This step is a little more tricky as you need a C/C++ style wrapper version of np.dot (so in this case, be sure the call to the dgemm function is with nogil: above it & indent the function the next line to release the gil with 4 space indent SO requires): https://gist.github.com/pv/5437087. That example looks to work (although you'll have to save the include f2pyptr.h file out and put it where your project is being built; I also suspect you should add cimport numpy as np); if not and it needs mods you can see as I've done in another post here: Calling BLAS / LAPACK directly using the SciPy interface and Cython (pointer issue?)/- also how to add MKL Then you'll need to add from cython.parallel cimport prange at the top and change all the loops to prange from range and be sure all your prange sections are nogil and all variables are cdef declared before operated on . In addition you'll have to add -openmp to your setup.py in the compiler arguments as well as a link to its include libs. Ask more questions if you need clarification. This isn't as easy as it should be but with a little guidance becomes quite simple. Basically once your setup.py is modified to include everything it will work going forward.
3) although probably easiest to fix - get rid of that List. Make it a numpy array or pandas dataframe if you need text and data. Whenever I've used lists for data the slowdown had been incredible.