On Fri, Jul 10, 2009 at 4:36 PM, John Schulman<[email protected]> wrote: > Wow! Thanks for the quick reply! > > When I run it with your changes, plus removing bounds checks, it goes > faster than when I build it with weave.inline and weave.ext_tools > > pure python 7.932142 > weave 0.084316 > ext_tools 0.081650 > cython 0.076264
Glad to see it do well! Although the credit goes to Cython as a whole & Dag Sverre Seljebotn for putting the buffer stuff in place last year. BTW, the reason it was slow in your original version is that it used Python C API PyObject_GetItem /SetItem in the inner loop, boxed the integers into Python Objects, and created tuples for the getitem/setitem calls. Adding the buffer declarations told Cython to generate fast indexing on the underlying ndarray data and not generate any Python API calls. One way you can tell is by looking at the generated C code (or run with cython -a cython_script.pyx for a html annotation file) and look inside the inner loop to see if there are any Python C API calls there. The first order of business (and often last) is to arrange things to get Python API calls removed. > > > On Fri, Jul 10, 2009 at 5:21 PM, Kurt Smith<[email protected]> wrote: >> On Fri, Jul 10, 2009 at 3:55 PM, John Schulman<[email protected]> wrote: >>> After reading the "Cython for Numpy Users" page, I tried rewriting a >>> function in cython, which I had previously implemented in python and >>> then scipy.weave. >>> It runs just as slow as the python version, which is 100 times slower >>> than the C version. >>> Maybe someone can tell me what is slowing it down. >>> Thanks, >>> John >>> >>> #cython_scripts.py >>> from __future__ import division >>> import numpy as np >>> cimport numpy as np >>> >>> DTYPE = np.int >>> ctypedef np.int_t DTYPE_t >>> >>> DTYPE2 = np.double >>> ctypedef np.double_t DTYPE2_t >>> >>> cdef inline int int_max(int a, int b): return a if a >= b else b >>> cdef inline int int_min(int a, int b): return a if a <= b else b >>> >>> def firstpass_labels(np.ndarray[DTYPE2_t,ndim=2] arr,list links,int >>> s_back,double thresh): >>> cdef int n_s,n_ch >>> n_s = arr.shape[0]; n_ch = arr.shape[1] >>> cdef np.ndarray labels = np.zeros([n_s,n_ch],dtype=DTYPE) >>> cdef DTYPE_t c_label >>> c_label = 1 >>> cdef int i_s, i_ch, j_s, j_ch, j_ch_ind, j_sstart >>> >>> cdef np.ndarray links_arr = np.zeros([n_ch,n_ch+1],dtype=DTYPE) >>> for (source,targs) in enumerate(links): >>> links_arr[source,0] = len(targs) >>> links_arr[source,1:(len(targs)+1)] = links[source] >>> >>> for i_s in range(n_s): >>> for i_ch in range(n_ch): >>> if arr[i_s,i_ch] > thresh: >>> j_sstart = int_max(0,i_s-s_back) >>> for j_s in range(j_sstart,i_s+1): >>> for j_ch_ind in range(1,links_arr[i_ch][0]+1): >>> j_ch = links_arr[i_ch,j_ch_ind] >>> if labels[j_s,j_ch] != 0: >>> labels[i_s,i_ch] = labels[j_s,j_ch] >>> >>> if labels[i_s,i_ch] == 0: >>> labels[i_s,i_ch] = c_label >>> c_label += 1 >>> >>> return labels >>> >>> >>> >>> And here's the function I use to run it: >>> #test_cy.py >>> import time >>> import numpy as np >>> >>> import pyximport; pyximport.install() >>> import cython_scripts >>> >>> >>> >>> arr = np.random.random( (400000,3)) >>> s_back = 3 >>> thresh = .7 >>> links2 = [[0,1],[0,1,2],[1,2]] >>> t = time.time() >>> print cython_scripts.firstpass_labels(arr,links2,s_back,thresh) >>> print "cython %f"%(time.time()-t) >> >> I get a X 70-80 speedup with the following changes. You can boost it >> more by turning off boundschecking & declaring every buffer's mode to >> be 'c'; see the documentation here: >> >> http://docs.cython.org/docs/numpy_tutorial.html#tuning-indexing-further >> >> #--------------------------------------------------------------------------- >> from __future__ import division >> import numpy as np >> cimport numpy as np >> >> DTYPE = np.int >> ctypedef np.int_t DTYPE_t >> >> DTYPE2 = np.double >> ctypedef np.double_t DTYPE2_t >> >> cdef inline int int_max(int a, int b): return a if a >= b else b >> cdef inline int int_min(int a, int b): return a if a <= b else b >> >> def firstpass_labels(np.ndarray[DTYPE2_t,ndim=2] arr,list links,int >> s_back,double thresh): >> cdef int n_s,n_ch >> n_s = arr.shape[0]; n_ch = arr.shape[1] >> # XXX: changed! >> cdef np.ndarray[DTYPE_t, ndim=2] labels = np.zeros([n_s,n_ch],dtype=DTYPE) >> cdef DTYPE_t c_label >> c_label = 1 >> cdef int i_s, i_ch, j_s, j_ch, j_ch_ind, j_sstart >> >> #XXX: changed! >> cdef np.ndarray[DTYPE_t, ndim=2] links_arr = >> np.zeros([n_ch,n_ch+1],dtype=DTYPE) >> for (source,targs) in enumerate(links): >> links_arr[source,0] = len(targs) >> links_arr[source,1:(len(targs)+1)] = links[source] >> >> for i_s in range(n_s): >> for i_ch in range(n_ch): >> if arr[i_s,i_ch] > thresh: >> j_sstart = int_max(0,i_s-s_back) >> for j_s in range(j_sstart,i_s+1): >> for j_ch_ind in range(1,links_arr[i_ch,0]+1): >> j_ch = links_arr[i_ch,j_ch_ind] >> if labels[j_s,j_ch] != 0: >> labels[i_s,i_ch] = labels[j_s,j_ch] >> >> if labels[i_s,i_ch] == 0: >> labels[i_s,i_ch] = c_label >> c_label += 1 >> >> return labels >> #------------------------------------------------------------------------------------------------------------------- >> _______________________________________________ >> Cython-dev mailing list >> [email protected] >> http://codespeak.net/mailman/listinfo/cython-dev >> > _______________________________________________ > Cython-dev mailing list > [email protected] > http://codespeak.net/mailman/listinfo/cython-dev > _______________________________________________ Cython-dev mailing list [email protected] http://codespeak.net/mailman/listinfo/cython-dev
