py_find_1st is a numpy extension that allows to find the first index into an 1D-array that validates a boolean condition that can consist of a comparison operator and a limit value.
This extension solves the very frequent problem of finding first indices without requiring to read the full array.
The call sequence
import numpy as np
import utils_find_1st as utf1st
limit = 0.
rr= np.random.randn(100)
ind = utf1st.find_1st(rr < limit, True, utf1st.cmp_equal)
and more efficiently
ind = utf1st.find_1st(rr, limit, utf1st.cmp_smaller)
is equivalent to
import numpy as np
limit = 0.
rr= np.random.randn(100)
ind = np.flatnonzero(rr < limit)
if len(ind) :
ret = ind[0]
else:
ret = -1
py_find_1st is written as a numpy extension making use of a templated implementation of the find_1st function that currently supports operating on arrays of dtypes:
[np.float64, np.float32, np.int64, np.int32, np.bool]
Comparison operators are selected using integer opcodes with the following meaning:
opcode == utils_find_1st.cmp_smaller -> comp op: <
opcode == utils_find_1st.cmp_smaller_eq -> comp op: <=
opcode == utils_find_1st.cmp_equal -> comp op: ==
opcode == utils_find_1st.cmp_not_equal -> comp op: !=
opcode == utils_find_1st.cmp_larger -> comp op: <
opcode == utils_find_1st.cmp_larger_eq -> comp op: <=
The runtime difference is strongly depending on the number of true cases in the array. If the condition is never valid runtime is the same - both implementations do not produce a valid index and need to compare the full array - but on case that there are matches np.flatnonzero needs to run through the full array and needs to create a result array with size that depends o the number of matches while find_1st only produces a scalar result and only needs to compare the array until the first match is found.
Depending on the size of the array and the number of matches the speed difference can be very significant (easily > factor 10)
run test/test_find_1st.py which should display "all tests passed!"
We can easily compare the runtime using the three lines
In [6]: timeit ind = np.flatnonzero(rr < limit)[0]
1.69 $\mu$s $\pm$ 24.5 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
In [4]: timeit ind = utf1st.find_1st(rr < limit, True, utf1st.cmp_equal)
1.13 $\mu$s $\pm$ 18.9 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
In [5]: timeit ind = utf1st.find_1st(rr, limit, utf1st.cmp_smaller)
270 ns $\pm$ 5.57 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
Which shows the rather significant improvement obtained by the last version that does not require to perform all comparisons of the 100 elements. In the above case the second element is tested positive. In the worst case, where no valid element is present all comparisons have to be performed and flatnonzero does not need to create a results array, and therefore performance should be similar. For the small array sizes we used so far the overhead of np.flanonzero is dominating the costs as can be seen in the following.
In [9]: limit = -1000.
In [10]: timeit ind = np.flatnonzero(rr < limit)
1.56 $\mu$s $\pm$ 13.8 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
In [11]: timeit ind = utf1st.find_1st(rr<limit, True, utf1st.cmp_equal)
1.16 $\mu$s $\pm$ 7.07 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
In [12]: timeit ind = utf1st.find_1st(rr, limit, utf1st.cmp_smaller)
314 ns $\pm$ 3.36 ns per loop (mean $\pm$ std. dev. of 7 runs, 1000000 loops each)
For a significantly larger array size costs become more comparable
rr= np.random.randn(10000)
In [13]: timeit ind = np.flatnonzero(rr < limit)
4.87 $\mu$s $\pm$ 101 ns per loop (mean $\pm$ std. dev. of 7 runs, 100000 loops each)
In [14]: timeit ind = utf1st.find_1st(rr<limit, True, utf1st.cmp_equal)
8.95 $\mu$s $\pm$ 497 ns per loop (mean $\pm$ std. dev. of 7 runs, 100000 loops each)
In [15]: timeit ind = utf1st.find_1st(rr, limit, utf1st.cmp_smaller)
4.4 $\mu$s $\pm$ 47.9 ns per loop (mean $\pm$ std. dev. of 7 runs, 100000 loops each)
Which demonstrates that even in this case the find_1st extension is more efficient besides if the boolean intermediate array is used in line 14.
This result is a bit astonishing as the overhead involved in passing the boolean intermediate array into the find_1st extension seems rather large compared to the simple boolean comparison
In [35]: timeit ind = rr < limit
3.31 $\mu$s $\pm$ 47.3 ns per loop (mean $\pm$ std. dev. of 7 runs, 100000 loops each)
The clarification of this remaining issue needs further investigation. Any comments are welcome.
- fixed test script for numpy > 1.20
- fixed Makefile under macosx ARM64
- fixed problems with numpy dependency handling (thanks to xmatthias). Now use oldest-supported-numpy instead of using the most recent numpy.
- added support for automatic installation of requirements
- add and support pre-release tags in the version number
- use hashlib to calculate the README checksum.
- support testing via
make check
- Removed setting stdlib for clang in setup.py - the default should do just fine.
- Removed ez_setup.py that seems to be no longer maintained by setuptools maintainers.
- Use NPY_INT64/NPY_INT32 instead of NPY_INT/NPY_LONG such that the test does not rely on the compiler specific int sizes.
- fixed bug in cmp operator values that were not coherent on the python and C++ side
- support arbitrary strides for one dimensional arrays
- Added test script
- Changed compiler test to hopefully work for MSVC under windows.
- Removed more non ascii elements in README.
- Fixed non ascii elements in README that led to problems with some python configurations.
- Fixed setup.py problems: on the fly generation of LONG_DESCRIPTION file.
- Moved to github
- Force using c++ compiler
- initial release
Copyright (C) 2017 IRCAM
GPL see file Copying.txt
Axel Roebel