Python Line_profiler and Cython function

Question

So I'm trying to profile a function within a python script of my own using line_profiler, because I want line-by-line timings. The only problem is that the function is a Cython one, and line_profiler isn't working correctly. On the first runs it was just crashing with an error. I then added

!python
cython: profile=True
cython: linetrace=True
cython: binding=True

at the top of my script and now it runs fine, except the timings and statistics are blank!

Is there a way to use line_profiler with a Cythonized function?

I could profile the non-Cythonized function, but it's so much slower than the Cythonized one that I could not use the information coming from the profiling - the slowness of the pure python one would make it impossible how I could improve the Cython one.

Here is the code of the function I'd want to profile:

class motif_hit(object):
__slots__ = ['position', 'strand']

def __init__(self, int position=0, int strand=0):
    self.position = position
    self.strand = strand

#the decorator for line_profiler
@profile
def find_motifs_cython(list bed_list, list matrices=None, int limit=0, int mut=0):
    cdef int q = 3
    cdef list bg = [0.25, 0.25, 0.25, 0.25]
    cdef int matrices_length = len(matrices)
    cdef int results_length = 0
    cdef int results_length_shuffled = 0
    cdef np.ndarray upper_adjust_list = np.zeros(matrices_length, np.int)
    cdef np.ndarray lower_adjust_list = np.zeros(matrices_length, np.int)
    #this one need to be a list for MOODS
    cdef list threshold_list = [None for _ in xrange(matrices_length)]
    cdef list matrix_list = [None for _ in xrange(matrices_length)]
    cdef np.ndarray results_list = np.zeros(matrices_length, np.object)
    cdef int count_seq = len(bed_list)
    cdef int mat
    cdef int i, j, k
    cdef int position, strand
    cdef list result, results, results_shuffled
    cdef dict result_temp
    cdef int length
    if count_seq > 0:
        for mat in xrange(matrices_length):
            matrix_list[mat] = matrices[mat]['matrix'].tolist()
            #change that for a class
            results_list[mat] = {'kmer': matrices[mat]['kmer'],
                                 'motif_count': 0,
                                 'pos_seq_count': 0,
                                 'motif_count_shuffled': 0,
                                 'pos_seq_count_shuffled': 0,
                                 'ratio': 0,
                                 'sequence_positions': np.empty(count_seq, np.object)}
            length = len(matrices[mat]['kmer'])
            #wrong with imbalanced matrices
            upper_adjust_list[mat] = int(ceil(length / 2.0))
            lower_adjust_list[mat] = int(floor(length / 2.0))
            #upper_adjust_list[mat] = 0
            #lower_adjust_list[mat] = 0
            #-0.1 to adjust for a division floating point bug (4.99999 !< 5, but is < 4.9!)
            threshold_list[mat] = MOODS.max_score(matrix_list[mat]) - float(mut) - 0.1

        #for each sequence
        for i in xrange(count_seq):
            item = bed_list[i]
            #TODO: remove the Ns, but it might unbalance
            results = MOODS.search(str(item.sequence[limit:item.total_length - limit]), matrix_list, threshold_list, q=q, bg=bg, absolute_threshold=True, both_strands=True)
            results_shuffled = MOODS.search(str(item.sequence_shuffled[limit:item.total_length - limit]), matrix_list, threshold_list, q=q, bg=bg, absolute_threshold=True, both_strands=True)
            results = results[0:len(matrix_list)]
            results_shuffled = results_shuffled[0:len(matrix_list)]
            results_length = len(results)
            #for each matrix
            for j in xrange(results_length):
                result = results[j]
                result_shuffled = results_shuffled[j]
                upper_adjust = upper_adjust_list[j]
                lower_adjust = lower_adjust_list[j]
                result_length = len(result)
                result_length_shuffled = len(result_shuffled)
                if result_length > 0:
                    results_list[j]['pos_seq_count'] += 1
                    results_list[j]['sequence_positions'][i] = np.empty(result_length, np.object)
                    #for each motif
                    for k in xrange(result_length):
                        position = result[k][0]
                        strand = result[k][1]
                        if position >= 0:
                                strand = 0
                                adjust = upper_adjust
                        else:
                                position = -position
                                strand = 1
                                adjust = lower_adjust
                        results_list[j]['motif_count'] += 1
                        results_list[j]['sequence_positions'][i][k] = motif_hit(position + adjust + limit, strand)

                if result_length_shuffled > 0:
                    results_list[j]['pos_seq_count_shuffled'] += 1
                    #for each motif
                    for k in xrange(result_length_shuffled):
                        results_list[j]['motif_count_shuffled'] += 1

                #j = j + 1
            #i = i + 1

        for i in xrange(results_length):
            result_temp = results_list[i]
            result_temp['ratio'] = float(result_temp['pos_seq_count']) / float(count_seq)
    return results_list

I'm pretty sure the triple nested loop is the main slow part - it's job is just to rearrange the results coming from MOODS, the C module doing the main work.

Answer 1

Till Hoffmann has useful information on using line_profiler with Cython here: How to profile cython functions line-by-line.

I quote his solution:

Robert Bradshaw helped me to get Robert Kern's line_profiler tool working for cdef functions and I thought I'd share the results on stackoverflow.

In short, set up a regular .pyx file and build script and pass to cythonize the linetrace compiler directive to enable both profiling and line tracing:

from Cython.Build import cythonize

cythonize('hello.pyx', compiler_directives={'linetrace': True})

You may also want to set the (undocumented) directive binding to True.

Also, you should define the C macro CYTHON_TRACE=1 by modifying your extensions setup such that

extensions = [
    Extension('test', ['test.pyx'], define_macros=[('CYTHON_TRACE', '1')])
]

A working example using the %%cython magic in the iPython notebook is here: http://nbviewer.ipython.org/gist/tillahoffmann/296501acea231cbdf5e7

Answer 2

Api was changed. Now:

from Cython.Compiler.Options import get_directive_defaults
directive_defaults = get_directive_defaults()
directive_defaults['linetrace'] = True
directive_defaults['binding'] = True