Some NumPy ufuncs, and some related tools.
ufunclab uses meson-python as its build system. Currently no
pre-built wheels are provided, so to install it, you'll have to check
out the git repository, and in the top-level directory, run
pip install .
in the terminal or command prompt. To run the unit tests after installing, run
pytest --pyargs ufunclab
To build ufunclab, a C99-compatible C compiler and a C++14-compatible C++
compiler are required.
The unit tests require pytest.
The test suite is run with Python versions 3.9 to 3.11, with several
recent releases of NumPy, and with the NumPy main development branch.
ufunclab might work with older Python and NumPy versions, but they
are not regularly tested.
Links to reference material related to NumPy's C API for ufuncs and gufuncs are given below.
Element-wise ufuncs
Most of the element-wise ufuncs are implemented by writing the core
calculation as a templated C++ function, and using some Python code to
automate the generation of all the necessary boilerplate and wrappers
that implement a ufunc around the core calculation. The exceptions
are logfactorial, log1p, loggamma1p, issnan, and cabssq, which
are implemented in C, with all boilerplate code written "by hand" in the
C file.
| Function | Description |
|---|---|
logfactorial |
Log of the factorial of integers |
issnan |
Like isnan, but for signaling nans only. |
next_less |
Equivalent to np.nextafter(x, -inf) |
next_greater |
Equivalent to np.nextafter(x, inf) |
abs_squared |
Squared absolute value |
cabssq |
Squared absolute value for complex input only |
deadzone |
Deadzone function |
step |
Step function |
linearstep |
Piecewise linear step function |
smoothstep3 |
Smooth step using a cubic polynomial |
invsmoothstep3 |
Inverse of smoothstep3 |
smoothstep5 |
Smooth step using a degree 5 polynomial |
trapezoid_pulse |
Trapezoid pulse function |
pow1pm1 |
Compute (1 + x)**y - 1 |
expint1 |
Exponential integral E₁ for real inputs |
log1p_doubledouble |
log(1 + z) for complex z. |
log1p_theorem4 |
log(1 + z) for complex z. |
logexpint1 |
Logarithm of the exponential integral E₁ |
loggamma1p |
Logarithm of gamma(1 + x) for real x > -1. |
logistic |
The standard logistic sigmoid function |
logistic_deriv |
Derivative of the standard logistic sigmoid function |
log_logistic |
Logarithm of the standard logistic sigmoid function |
swish |
The 'swish' function--a smoothed ramp |
hyperbolic_ramp |
A smoothed ramp whose graph is a hyperbola |
exponential_ramp |
A smoothed ramp with exponential convergence to asymptotes |
yeo_johnson |
Yeo-Johnson transformation |
inv_yeo_johnson |
Inverse of the Yeo-Johnson transformation |
erfcx |
Scaled complementary error function |
| Normal distribution functions | Functions for the normal distribution: cdf, sf, logcdf, logsf |
| Semivariograms | Several semivariograms used in kriging interpolation |
Generalized ufuncs
Note, for anyone looking at the source code: some of these implementations
are in C and use the legacy NumPy templating language (look for filenames
that end in .src); others use templated C++ functions combined with code
generation tools that can be found in tools/cxxgen. The .src files are
processed with the script in ufunclab/tools/conv_template.py.
| Function | Description |
|---|---|
first |
First value that matches a target comparison |
argfirst |
Index of the first occurrence of a target comparison |
argmin |
Like numpy.argmin, but a gufunc |
argmax |
Like numpy.argmax, but a gufunc |
minmax |
Minimum and maximum |
argminmax |
Indices of the min and the max |
min_argmin |
Minimum value and its index |
max_argmax |
Maximum value and its index |
searchsortedl |
Find position for given element in sorted seq. |
searchsortedr |
Find position for given element in sorted seq. |
peaktopeak |
Alternative to numpy.ptp |
all_same |
Check all values are the same |
gmean |
Geometric mean |
hmean |
Harmonic mean |
meanvar |
Mean and variance |
mad |
Mean absolute difference (MAD) |
rmad |
Relative mean absolute difference (RMAD) |
gini |
Gini coefficient |
rms |
Root-mean-square for real and complex inputs |
vnorm |
Vector norm |
vdot |
Vector dot product for real floating point arrays |
pearson_corr |
Pearson's product-moment correlation coefficient |
wjaccard |
Weighted Jaccard index. |
cross2 |
2-d vector cross product (returns scalar) |
cross3 |
3-d vector cross product |
tri_area |
Area of triangles in n-dimensional space |
fillnan1d |
Replace nan using linear interpolation |
linear_interp1d |
Linear interpolation, like numpy.interp |
backlash |
Backlash operator |
hysteresis_relay |
Relay with hysteresis (Schmitt trigger) |
sosfilter |
SOS (second order sections) linear filter |
sosfilter_ic |
SOS linear filter with initial condition |
sosfilter_ic_contig |
SOS linear filter with contiguous array inputs |
multivariate_logbeta |
Logarithm of the multivariate beta function |
Other tools
| Function | Description |
|---|---|
gendot |
Create a new gufunc that composes two ufuncs |
ufunc_inspector |
Display ufunc information |
logfactorial is a ufunc that computes the natural logarithm of the
factorial of the nonnegative integer x. (nan is returned for negative
input.)
For example,
>>> from ufunclab import logfactorial
>>> logfactorial([1, 10, 100, 1000])
array([ 0. , 15.10441257, 363.73937556, 5912.12817849])
issnan is an element-wise ufunc with a single input that acts like
the standard isnan function, but it returns True only for
signaling nans.
The current implementation only handles the floating point types np.float16,
np.float32 and np.float64.
>>> import numpy as np
>>> from ufunclab import issnan
>>> x = np.array([12.5, 0.0, np.inf, 999.0, np.nan], dtype=np.float32)
Put a signaling nan in x[1]. (The nan in x[4] is a quiet nan, and
we'll leave it that way.)
>>> v = x.view(np.uint32)
>>> v[1] = 0b0111_1111_1000_0000_0000_0000_0000_0011
>>> x
array([ 12.5, nan, inf, 999. , nan], dtype=float32)
>>> np.isnan(x)
array([False, True, False, False, True])
Note that NumPy displays both quiet and signaling nans as just nan,
and np.isnan(x) returns True for both quiet and signaling nans (as
it should).
issnan(x) indicates which values are signaling nans:
>>> issnan(x)
array([False, True, False, False, False])
next_less is an element-wise ufunc with a single input that
is equivalent to np.nextafter with the second argument set to -inf.
>>> import numpy as np
>>> from ufunclab import next_less
>>> next_less(np.array([-12.5, 0, 1, 1000], dtype=np.float32))
array([-1.2500001e+01, -1.4012985e-45, 9.9999994e-01, 9.9999994e+02],
dtype=float32)
next_greater is an element-wise ufunc with a single input that
is equivalent to np.nextafter with the second argument set to inf.
>>> import numpy as np
>>> from ufunclab import next_greater
>>> next_greater(np.array([-12.5, 0, 1, 1000], dtype=np.float32))
array([-1.24999990e+01, 1.40129846e-45, 1.00000012e+00, 1.00000006e+03],
dtype=float32)
abs_squared(z) computes the squared absolute value of z.
This is an element-wise ufunc with types 'f->f', 'd->d',
'g->g', 'F->f', 'D->d', and 'G->g'. For real input,
the result is z**2. For complex input, it is
z.real**2 + z.imag**2, which is equivalent to z*conj(z),
where conj(z) is the complex conjugate of z.
>>> import numpy as np
>>> from ufunclab import abs_squared
>>> abs_squared.types
['f->f', 'd->d', 'g->g', 'F->f', 'D->d', 'G->g']
>>> x = np.array([-1.5, 3.0, 9.0, -10.0], dtype=np.float32)
>>> abs_squared(x)
array([ 2.25, 9. , 81. , 100. ], dtype=float32)
>>> z = np.array([-3+4j, -1, 1j, 13, 0.5-1.5j])
>>> abs_squared(z)
array([ 25. , 1. , 1. , 169. , 2.5])
cabssq(z) computes the squared absolute value of z for complex input only.
This is the same calculation as abs_squared, but the implementation is
different. cabssq is implemented in C with the inner loop functions
implemented "by hand", with no C++ or NumPy templating. cabssq is generally
faster than abs_squared, because it avoids some of the overhead that occurs
in the code generated in the implementation of abs_squared, and it allows
the compiler to optimize the code more effectively.
deadzone(x, low, high) is a ufunc with three inputs and one output.
It computes the "deadzone" response of a signal:
{ 0 if low <= x <= high
f(x) = { x - low if x < low
{ x - high if x > high
The function is similar to the deadzone block of Matlab's Simulink library. The function is also known as a soft threshold.
Here's a plot of deadzone(x, -0.25, 0.5):
The script deadzone_demo2.py in the examples directory generates
the plot
The ufunc step(x, a, flow, fa, fhigh) returns flow for
x < a, fhigh for x > a, and fa for x = a.
The Heaviside function can be implemented as step(x, 0, 0, 0.5, 1).
The script step_demo.py in the examples directory generates
the plot
The ufunc linearstep(x, a, b, fa, fb) returns fa for
x <= a, fb for x >= b, and uses linear interpolation
from fa to fb in the interval a < x < b.
The script linearstep_demo.py in the examples directory generates
the plot
The ufunc smoothstep3(x, a, b, fa, fb) returns fa for
x <= a, fb for x >= b, and uses a cubic polynomial in
the interval a < x < b to smoothly transition from fa to fb.
The script smoothstep3_demo.py in the examples directory generates
the plot
The ufunc invsmoothstep3(y, a, b, fa, fb) is the inverse of
smoothstep3(x, a, b, fa, fb).
The script invsmoothstep3_demo.py in the examples directory generates
the plot
The function smoothstep5(x, a, b, fa, fb) returns fa for
x <= a, fb for x >= b, and uses a degree 5 polynomial in
the interval a < x < b to smoothly transition from fa to fb.
The script smoothstep5_demo.py in the examples directory generates
the plot
trapezoid_pulse(x, a, b, c, d, amp) is a ufunc that computes
a trapezoid pulse. The function is 0 for x <= a or x >= d,
amp for b <= x <= c, and a linear ramp in the intervals
[a, b] and [c, d].
Here's a plot of trapezoid_pulse(x, 1, 3, 4, 5, 2) (generated by
the script trapezoid_pulse_demo.py in the examples directory):
pow1pm1(x, y) computes (1 + x)**y - 1 for x >= -1.
The calculation is formulated to avoid loss of precision when
(1 + x)**y is close to 1.
>>> from ufunclab import pow1pm1
The following result provides full machine precision:
>>> x = -0.125
>>> y = 3.25e-12
>>> pow1pm1(x, y)
-4.3397702602960437e-13
The naive calculation provides less than six digits of precision:
>>> (1 + x)**y - 1
-4.339861803259737e-13
expint1(x) computes the exponential integral E₁ for the real input x.
>>> from ufunclab import expint1
>>> expint1([0.25, 2.5, 25])
array([1.04428263e+00, 2.49149179e-02, 5.34889976e-13])
log1p_doubledouble(z) computes log(1 + z) for complex z. This is an alternative
to numpy.log1p and scipy.special.log1p.
The function uses double-double numbers for some intermediate calculations to avoid loss of precision near the circle |z + 1| = 1 in the complex plane.
>>> from ufunclab import log1p_doubledouble
>>> z = -0.57113-0.90337j
>>> log1p_doubledouble(z)
(3.4168883248419116e-06-1.1275564209486122j)
Currently, the only "inner loop" implemented for this ufunc is for
the data type np.complex128, so it will always return a complex result:
>>> log1p_doubledouble.types
['D->D']
>>> log1p_doubledouble([-6e-3, 0, 1e-12, 5e-8, 0.25, 1])
array([-6.01807233e-03+0.j, 0.00000000e+00+0.j, 1.00000000e-12+0.j,
4.99999988e-08+0.j, 2.23143551e-01+0.j, 6.93147181e-01+0.j])
log1p_theorem4(z) computes log(1 + z) for complex z. This is an alternative
to numpy.log1p and scipy.special.log1p.
The function uses the "trick" given as Theorem 4 of Goldberg's paper "What every computer scientist should know about floating-point arithmetic".
The precision provided by this function depends on the precision of the function
clog (complex logarithm) provided by the underlying C math library that is used
to build ufunclab.
>>> from ufunclab import log1p_theorem4
>>> z = -0.57113-0.90337j
>>> log1p_theorem4(z)
(3.4168883248419116e-06-1.1275564209486122j)
Currently, the only "inner loop" implemented for this ufunc is for
the data type np.complex128, so it will always return a complex result:
>>> log1p_theorem4.types
['D->D']
>>> log1p_theorem4([-6e-3, 0, 1e-12, 5e-8, 0.25, 1])
array([-6.01807233e-03+0.j, 0.00000000e+00+0.j, 1.00000000e-12+0.j,
4.99999988e-08+0.j, 2.23143551e-01+0.j, 6.93147181e-01+0.j])
logexpint1(x) computes the logarithm of the exponential integral E₁ for the real input x.
expint1(x) underflows to 0 for sufficiently large x:
>>> from ufunclab import expint1, logexpint1
>>> expint1([650, 700, 750, 800])
array([7.85247922e-286, 1.40651877e-307, 0.00000000e+000, 0.00000000e+000])
logexpint1 avoids the underflow by computing the logarithm of the value:
>>> logexpint1([650, 700, 750, 800])
array([-656.47850729, -706.55250586, -756.62140388, -806.68585939])
loggamma1p computes log(gamma(1 + x)) for real x > -1.
It avoids the loss of precision in the expression 1 + x that occurs
when x is very small.
>>> from ufunclab import loggamma1p
>>> x = -3e-11
>>> loggamma1p(x)
1.7316469947786207e-11
That result is accurate to machine precision. The naive calculation
loses precision in the sum 1 + x; in the following result that uses
math.lgamma, only about five digits are correct:
>>> import math
>>> math.lgamma(1 + x)
1.7316037492776104e-11
logistic(x) computes the standard logistic sigmoid function.
The script logistic_demo.py in the examples directory generates
this plot:
logistic_deriv(x) computes the derivative of the standard logistic sigmoid
function.
See logistic (above) for a plot.
log_logistic(x) computes the logarithm of the standard logistic sigmoid
function.
>>> import numpy as np
>>> from ufunclab import log_logistic
>>> x = np.array([-800, -500, -0.5, 10, 250, 500])
>>> log_logistic(x)
array([-8.00000000e+002, -5.00000000e+002, -9.74076984e-001,
-4.53988992e-005, -2.66919022e-109, -7.12457641e-218])
Compare that to the output of log(expit(x)), which triggers a warning
and loses all precision for inputs with large magnitudes:
>>> from scipy.special import expit
>>> np.log(expit(x))
<stdin>:1: RuntimeWarning: divide by zero encountered in log
array([ -inf, -5.00000000e+02, -9.74076984e-01,
-4.53988992e-05, 0.00000000e+00, 0.00000000e+00])
swish(x, beta) computes x * logistic(beta*x), where logistic(x)
is the standard logistic sigmoid function. The function is a type
of smoothed ramp.
hyperbolic_ramp(x, a) computes the function
hyperbolic_ramp(x, a) = (x + sqrt(x*x + 4*a*a))/2
It is a smoothed ramp function. The scaling of the parameters is chosen
so that hyperbolic_ramp(0, a) is a.
exponential_ramp(x, a) computes the function
exponential_ramp(x, a) = a*log_2(1 + 2**(x/a))
It is a smoothed ramp function that converges exponentially fast to
the asymptotes. The scaling of the parameters is chosen so that
exponential_ramp(0, a) is a.
The function is also known as the "softplus" function.
yeo_johnson computes the Yeo-Johnson transform.
>>> import numpy as np
>>> from ufunclab import yeo_johnson
>>> yeo_johnson([-1.5, -0.5, 2.8, 7, 7.1], 2.3)
array([-0.80114069, -0.38177502, 8.93596922, 51.4905317 , 52.99552905])
>>> yeo_johnson.outer([-1.5, -0.5, 2.8, 7, 7.1], [-2.5, 0.1, 2.3, 3])
array([[-13.50294123, -2.47514321, -0.80114069, -0.6 ],
[ -1.15561576, -0.61083954, -0.38177502, -0.33333333],
[ 0.38578977, 1.42821388, 8.93596922, 17.95733333],
[ 0.39779029, 2.31144413, 51.4905317 , 170.33333333],
[ 0.39785786, 2.32674755, 52.99552905, 176.81366667]])
inv_yeo_johnson computes the inverse of the Yeo-Johnson transform.
>>> import numpy as np
>>> from ufunclab import inv_yeo_johnson, yeo_johnson
>>> x = inv_yeo_johnson([-1.5, -0.5, 2.8, 7, 7.1], 2.5)
>>> x
array([-15. , -0.77777778, 1.29739671, 2.21268904,
2.22998502])
>>> yeo_johnson(x, 2.5)
array([-1.5, -0.5, 2.8, 7. , 7.1])
erfcx(x) computes the scaled complementary error function,
exp(x**2) * erfc(x). The function is implemented for NumPy types
float32, float64 and longdouble (also known as float128):
>>> from ufunclab import erfcx
>>> erfcx.types
['f->f', 'd->d', 'g->g']
This example is run on a platform where the longdouble type
corresponds to a float128 with 80 bits of precision:
>>> import numpy as np
>>> b = np.longdouble('1.25e2000')
>>> x = np.array([-40, -1.0, 0, 2.5, 3000.0, b])
>>> x
array([-4.00e+0001, -1.00e+0000, 0.00e+0000, 1.00e+0003, 1.25e+2000],
dtype=float128)
>>> erfcx(x)
array([1.48662366e+0695, 5.00898008e+0000, 1.00000000e+0000,
5.64189301e-0004, 4.51351667e-2001], dtype=float128)
The submodule normal defines several functions for the standard normal
probability distrbution.
normal.cdf(x) computes the cumulative distribution function of the standard
normal distribution.
normal.logcdf(x) computes the natural logarithm of the CDF of the standard
normal distribution.
normal.sf(x) computes the survival function of the standard normal distribution.
This function is also known as the complementary CDF, and is often abbreviated
as ccdf.
normal.logsf(x) computes the natural logarithm of the survival function of the
standard normal distribution.
The submodule ufunclab.semivar defines the ufuncs exponential, linear and
spherical. The script semivar_demo.py in the examples directory generates
the following plot.
semivar.exponential(h, nugget, sill, rng) computes the exponential semivariogram.
semivar.linear(h, nugget, sill, rng) computes the linear semivariogram.
semivar.spherical(h, nugget, sill, rng) computes the spherical semivariogram.
first is a gufunc with signature (i),(),(),()->() that returns the first
value that matches a given comparison. The function signature is
first(x, op, target, otherwise), where op is one of the values in
ufunclab.op that specifies the comparison to be made. otherwise is the
value to be returned if no value in x satisfies the given relation with
target.
Find the first nonzero value in a:
>>> import numpy as np
>>> from ufunclab import first, op
>>> a = np.array([0, 0, 0, 0, 0, -0.5, 0, 1, 0.1])
>>> first(a, op.NE, 0.0, 0.0)
-0.5
Find the first value in each row of b that is less than 0.
If there is no such value, return 0:
>>> b = np.array([[10, 23, -10, 0, -9],
... [18, 28, 42, 33, 71],
... [17, 29, 16, 14, -7]], dtype=np.int8)
...
>>> first(b, op.LT, 0, 0)
array([-10, 0, -7], dtype=int8)
The function stops at the first occurrence, so it will be faster when the condition is met at the beginning of the array than when the condition doesn't occur until near the end of the array.
In the following, the condition occurs in the last element of x, and
the result of the timeit call is 0.6 seconds.
>>> from timeit import timeit
>>> x = np.ones(100000)
>>> x[-] = 0
>>> first(x, op.LT, 1.0, np.nan)
0.0
>>> timeit('first(x, op.LT, 1.0, np.nan)', number=20000, globals=globals())
0.6052625320153311
The time is reduced to 0.05 seconds when the first occurrence of the
condition is in the first element of x.
>>> x[0] = -1.0
>>> first(x, op.LT, 1.0, np.nan)
-1.0
>>> timeit('first(x, op.LT, 1.0, np.nan)', number=20000, globals=globals())
0.049594153999350965
argfirst is a gufunc (signature (i),(),()->()) that finds the index of
the first true value of a comparison of an array with a target value. If no
value is found, -1 is return. Some examples follow.
>>> import numpy as np
>>> from ufunclab import argfirst, op
Find the index of the first occurrence of 0 in x:
>>> x = np.array([10, 35, 19, 0, -1, 24, 0])
>>> argfirst(x, op.EQ, 0)
3
Find the index of the first nonzero value in a:
>>> a = np.array([0, 0, 0, 0, 0, -0.5, 0, 1, 0.1])
>>> argfirst(a, op.NE, 0.0)
5
argfirst is a gufunc, so it can handle higher-dimensional
array arguments, and among its gufunc-related parameters is
axis. By default, the gufunc operates along the last axis.
For example, here we find the location of the first nonzero
element in each row of b:
>>> b = np.array([[0, 8, 0, 0], [0, 0, 0, 0], [0, 0, 9, 2]],
... dtype=np.uint8)
>>> b
array([[0, 8, 0, 0],
[0, 0, 0, 0],
[0, 0, 9, 2]])
>>> argfirst(b, op.NE, np.uint8(0))
array([ 1, -1, 2])
If we give the argument axis=0, we tell argfirst to
operate along the first axis, which in this case is the
columns:
>>> argfirst(b, op.NE, np.uint8(0), axis=0)
array([-1, 0, 2, 2])
argmin is a gufunc with signature (i)->() that is similar to numpy.argmin.
>>> from ufunclab import argmin
>>> x = np.array([[11, 10, 10, 23, 31],
... [19, 20, 21, 22, 22],
... [16, 15, 16, 14, 14]])
>>> argmin(x, axis=1) # same as argmin(x)
array([1, 0, 3])
>>> argmin(x, axis=0)
array([0, 0, 0, 2, 2])
argmax is a gufunc with signature (i)->() that is similar to numpy.argmax.
>>> from ufunclab import argmax
>>> x = np.array([[11, 10, 10, 23, 31],
... [19, 20, 21, 22, 22],
... [16, 15, 16, 14, 14]])
>>> argmax(x, axis=1) # same as argmax(x)
array([4, 3, 0])
>>> argmax(x, axis=0)
array([1, 1, 1, 0, 0])
minmax is a gufunc (signature (i)->(2)) that simultaneously computes
the minimum and maximum of a NumPy array.
The function handles the standard integer and floating point types, and
object arrays. The function will not accept complex arrays, nor arrays with
the data types datetime64 or timedelta64. Also, the function does not
implement any special handling of nan, so the behavior of this function
with arrays containing nan is undefined.
For an input with more than one dimension, minmax is applied to the
last axis. For example, if a has shape (L, M, N), then minmax(a) has
shape (L, M, 2).
>>> x = np.array([5, -10, -25, 99, 100, 10], dtype=np.int8)
>>> minmax(x)
array([-25, 100], dtype=int8)
>>> np.random.seed(12345)
>>> y = np.random.randint(-1000, 1000, size=(3, 3, 5)).astype(np.float32)
>>> y
array([[[-518., 509., 309., -871., 444.],
[ 449., -618., 381., -454., 565.],
[-231., 142., 393., 339., -346.]],
[[-895., 115., -241., 398., 232.],
[-118., -287., -733., 101., 674.],
[-919., 746., -834., -737., -957.]],
[[-769., -977., 53., -48., 463.],
[ 311., -299., -647., 883., -145.],
[-964., -424., -613., -236., 148.]]], dtype=float32)
>>> mm = minmax(y)
>>> mm
array([[[-871., 509.],
[-618., 565.],
[-346., 393.]],
[[-895., 398.],
[-733., 674.],
[-957., 746.]],
[[-977., 463.],
[-647., 883.],
[-964., 148.]]], dtype=float32)
>>> mm.shape
(3, 3, 2)
>>> z = np.array(['foo', 'xyz', 'bar', 'abc', 'def'], dtype=object)
>>> minmax(z)
array(['abc', 'xyz'], dtype=object)
>>> from fractions import Fraction
>>> f = np.array([Fraction(1, 3), Fraction(3, 5),
... Fraction(22, 7), Fraction(5, 2)], dtype=object)
>>> minmax(f)
array([Fraction(1, 3), Fraction(22, 7)], dtype=object)
argminmax is a gufunc (signature (i)->(2)) that simultaneously
computes the argmin and argmax of a NumPy array.
>>> y = np.array([[-518, 509, 309, -871, 444, 449, -618, 381],
... [-454, 565, -231, 142, 393, 339, -346, -895],
... [ 115, -241, 398, 232, -118, -287, -733, 101]],
... dtype=np.float32)
>>> argminmax(y)
array([[3, 1],
[7, 1],
[6, 2]])
>>> argminmax(y, axes=[0, 0])
array([[0, 2, 1, 0, 2, 2, 2, 1],
[2, 1, 2, 2, 0, 0, 1, 0]])
min_argmin is a gufunc (signature (i)->(),()) that returns both
the extreme value and the index of the extreme value.
>>> x = np.array([[ 1, 10, 18, 17, 11],
... [15, 11, 0, 4, 8],
... [10, 10, 12, 11, 11]])
>>> min_argmin(x, axis=1)
(array([ 1, 0, 10]), array([0, 2, 0]))
max_argmax is a gufunc (signature (i)->(),()) that returns both
the extreme value and the index of the extreme value.
>>> x = np.array([[ 1, 10, 18, 17, 11],
... [15, 11, 0, 4, 8],
... [10, 10, 12, 11, 11]])
>>> max_argmax(x, axis=1)
(array([18, 15, 12]), array([2, 0, 2]))
>>> from fractions import Fraction as F
>>> y = np.array([F(2, 3), F(3, 4), F(2, 7), F(2, 5)])
>>> max_argmax(y)
(Fraction(3, 4), 1)
searchsortedl is a gufunc with signature (i),()->(). The function
is equivalent to numpy.searchsorted with side='left', but as a gufunc,
it supports broadcasting of its arguments. (Note that searchsortedl
does not provide the sorter parameter.)
>>> import numpy as np
>>> from ufunclab import searchsortedl
>>> searchsortedl([1, 1, 2, 3, 5, 8, 13, 21], [1, 4, 15, 99])
array([0, 4, 7, 8])
>>> arr = np.array([[1, 1, 2, 3, 5, 8, 13, 21],
... [1, 1, 1, 1, 2, 2, 10, 10]])
>>> searchsortedl(arr, [7, 8])
array([5, 6])
>>> searchsortedl(arr, [[2], [5]])
array([[2, 4],
[4, 6]])
searchsortedr is a gufunc with signature (i),()->(). The function
is equivalent to numpy.searchsorted with side='right', but as a gufunc,
it supports broadcasting of its arguments. (Note that searchsortedr
does not provide the sorter parameter.)
>>> import numpy as np
>>> from ufunclab import searchsortedr
>>> searchsortedr([1, 1, 2, 3, 5, 8, 13, 21], [1, 4, 15, 99])
array([2, 4, 7, 8])
>>> arr = np.array([[1, 1, 2, 3, 5, 8, 13, 21],
... [1, 1, 1, 1, 2, 2, 10, 10]])
>>> searchsortedr(arr, [7, 8])
array([5, 6])
>>> searchsortedr(arr, [[2], [5]])
array([[3, 6],
[5, 6]])
peaktopeak is a gufunc (signature (i)->()) that computes the
peak-to-peak range of a NumPy array. It is like the ptp method
of a NumPy array, but when the input is signed integers, the output
is an unsigned integer with the same bit width.
The function handles the standard integer and floating point types,
datetime64, timedelta64, and object arrays. The function does not
accept complex arrays. Also, the function does not implement any special
handling of nan, so the behavior of this function with arrays containing
nan is undefined (i.e. it might not do what you want, and the behavior
might change in the next update of the software).
>>> x = np.array([85, 125, 0, -75, -50], dtype=np.int8)
>>> p = peaktopeak(x)
>>> p
200
>>> type(p)
numpy.uint8
Compare that to the ptp method, which returns a value with the
same data type as the input:
>>> q = x.ptp()
>>> q
-56
>>> type(q)
numpy.int8
f is an object array of Fractions and has shape (2, 4).
>>> from fractions import Fraction
>>> f = np.array([[Fraction(1, 3), Fraction(3, 5),
... Fraction(22, 7), Fraction(5, 2)],
... [Fraction(-2, 9), Fraction(1, 3),
... Fraction(2, 3), Fraction(5, 9)]], dtype=object)
>>> peaktopeak(x)
array([Fraction(59, 21), Fraction(8, 9)], dtype=object)
dates is an array of datetime64.
>>> dates = np.array([np.datetime64('2015-11-02T12:34:50'),
... np.datetime64('2016-03-01T16:00:00'),
... np.datetime64('2015-07-02T21:20:19'),
... np.datetime64('2016-05-01T19:25:00')])
>>> dates
array(['2015-11-02T12:34:50', '2016-03-01T16:00:00',
'2015-07-02T21:20:19', '2016-05-01T19:25:00'],
dtype='datetime64[s]')
>>> timespan = peaktopeak(dates)
>>> timespan
numpy.timedelta64(26258681,'s')
>>> timespan / np.timedelta64(1, 'D') # Convert to number of days.
303.9199189814815
Casting works when the out argument is an array with dtype timedelta64.
For example,
>>> out = np.empty((), dtype='timedelta64[D]')
>>> peaktopeak(dates, out=out)
array(303, dtype='timedelta64[D]')
all_same is a gufunc (signature (i)->()) that tests that all the
values in the array along the given axis are the same.
(Note: handling of datetime64, timedelta64 and complex data types
are not implemented yet.)
>>> x = np.array([[3, 2, 2, 3, 2, 2, 3, 1, 3],
... [1, 2, 2, 2, 2, 2, 3, 1, 1],
... [2, 3, 3, 1, 2, 3, 3, 1, 2]])
>>> all_same(x, axis=0)
array([False, False, False, False, True, False, True, True, False])
>>> all_same(x, axis=1)
array([False, False, False])
Object arrays are handled.
>>> a = np.array([[None, "foo", 99], [None, "bar", "abc"]])
>>> a
array([[None, 'foo', 99],
[None, 'bar', 'abc']], dtype=object)
>>> all_same(a, axis=0)
array([ True, False, False])
gmean is a gufunc (signature (i)->()) that computes the
geometric mean.
For example,
>>> import numpy as np
>>> from ufunclab import gmean
>>> x = np.array([1, 2, 3, 5, 8], dtype=np.uint8)
>>> gmean(x)
2.992555739477689
>>> y = np.arange(1, 16).reshape(3, 5)
>>> y
array([[ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10],
[11, 12, 13, 14, 15]])
>>> gmean(y, axis=1)
array([ 2.60517108, 7.87256685, 12.92252305])
hmean is a gufunc (signature (i)->()) that computes the
harmonic mean.
For example,
>>> import numpy as np
>>> from ufunclab import hmean
>>> x = np.array([1, 2, 3, 5, 8], dtype=np.uint8)
>>> hmean(x)
2.316602316602317
>>> y = np.arange(1, 16).reshape(3, 5)
>>> y
array([[ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10],
[11, 12, 13, 14, 15]])
>>> hmean(y, axis=1)
array([ 2.18978102, 7.74431469, 12.84486077])
meanvar is a gufunc (signature (n),()->(2)) that computes both
the mean and variance in one function call.
For example,
>>> import numpy as np
>>> from ufunclab import meanvar
>>> meanvar([1, 2, 4, 5], 0) # Use ddof=0.
array([3. , 2.5])
Apply meanvar with ddof=1 to the rows of a 2-d array.
The output has shape (4, 2); the first column holds the
means, and the second column holds the variances.
>>> x = np.array([[1, 4, 4, 2, 1, 1, 2, 7],
... [0, 0, 9, 4, 1, 0, 0, 1],
... [8, 3, 3, 3, 3, 3, 3, 3],
... [5, 5, 5, 5, 5, 5, 5, 5]])
>>> meanvar(x, 1) # Use ddof=1.
array([[ 2.75 , 4.5 ],
[ 1.875, 10.125],
[ 3.625, 3.125],
[ 5. , 0. ]])
Compare to the results of numpy.mean and numpy.var:
>>> np.mean(x, axis=1)
array([2.75 , 1.875, 3.625, 5. ])
>>> np.var(x, ddof=1, axis=1)
array([ 4.5 , 10.125, 3.125, 0. ])
mad(x, unbiased) computes the mean absolute difference
of a 1-d array (gufunc signature is (n),()->()). When the second parameter
is False, mad is the standard calculation (sum of the absolute differences
divided by n**2). When the second parameter is True, mad is the unbiased
estimator (sum of the absolute differences divided by n*(n-1)).
For example,
>>> import numpy as np
>>> from ufunclab import mad
>>> x = np.array([1.0, 1.0, 2.0, 3.0, 5.0, 8.0])
>>> mad(x, False)
2.6666666666666665
>>> y = np.linspace(0, 1, 21).reshape(3, 7)**2
>>> y
array([[0. , 0.0025, 0.01 , 0.0225, 0.04 , 0.0625, 0.09 ],
[0.1225, 0.16 , 0.2025, 0.25 , 0.3025, 0.36 , 0.4225],
[0.49 , 0.5625, 0.64 , 0.7225, 0.81 , 0.9025, 1. ]])
>>> mad(y, False, axis=1)
array([0.03428571, 0.11428571, 0.19428571])
When the second parameter is True, the calculation is the unbiased
estimate of the mean absolute difference.
>>> mad(x, True)
3.2
>>> mad(y, True, axis=1)
array([0.04 , 0.13333333, 0.22666667])
rmad(x, unbiased) computes the relative mean absolute difference (gufunc
signature is (i),()->()).
rmad is twice the Gini coefficient.
For example,
>>> import numpy as np
>>> from ufunclab import rmad
>>> x = np.array([1.0, 1.0, 2.0, 3.0, 5.0, 8.0])
>>> rmad(x, False)
0.7999999999999999
>>> y = np.linspace(0, 1, 21).reshape(3, 7)**2
>>> y
array([[0. , 0.0025, 0.01 , 0.0225, 0.04 , 0.0625, 0.09 ],
[0.1225, 0.16 , 0.2025, 0.25 , 0.3025, 0.36 , 0.4225],
[0.49 , 0.5625, 0.64 , 0.7225, 0.81 , 0.9025, 1. ]])
>>> rmad(y, False, axis=1)
array([1.05494505, 0.43956044, 0.26523647])
When the second parameter is True, the calculation is based on the
unbiased estimate of the mean absolute difference (MAD).
>>> rmad(x, True)
0.96
>>> rmad(y, True, axis=1)
array([1.23076923, 0.51282051, 0.30944255])
gini(x, unbiased) is a gufunc with signature (n),()->() that computes the
Gini coefficient of the data in x.
>>> from ufunclab import gini
>>> gini([1, 2, 3, 4], False)
0.25
>>> income = [20, 30, 40, 50, 60, 70, 80, 90, 120, 150]
>>> gini(income, False)
0.3028169014084507
When the second parameter is True, the calculation is based on the
unbiased estimate of the mean absolute difference (MAD).
>>> gini([1, 2, 3, 4], True)
0.33333333333333337
>>> income = [20, 30, 40, 50, 60, 70, 80, 90, 120, 150]
>>> gini(income, True)
0.3364632237871674
rms(x) computes the root-mean-square value for a collection of values.
It is a gufunc with signature (n)->(). The implementation is for
float and complex types; integer types are cast to float.
>>> import numpy as np
>>> from ufunclab import rms
>>> x = np.array([1, 2, -1, 0, 3, 2, -1, 0, 1])
>>> rms(x)
1.5275252316519468
Compare to:
>>> np.sqrt(np.mean(x**2))
1.5275252316519468
A complex example:
>>> z = np.array([1-1j, 2+1.5j, -3-2j, 0.5+1j, 2.5j], dtype=np.complex64)
>>> rms(z)
2.3979158
An equivalent NumPy expression:
>>> np.sqrt(np.mean(z.real**2 + z.imag**2))
2.3979158
vnorm(x, p) computes the vector p-norm of 1D arrays. It is a gufunc with
signatue (i), () -> ().
For example, to compute the 2-norm of [3, 4]:
>>> import numpy as np
>>> from ufunclab import vnorm
>>> vnorm([3, 4], 2)
5.0
Compute the p-norm of [3, 4] for several values of p:
>>> vnorm([3, 4], [1, 2, 3, np.inf])
array([7. , 5. , 4.49794145, 4. ])
Compute the 2-norm of four 2-d vectors:
>>> vnorm([[3, 4], [5, 12], [0, 1], [1, 1]], 2)
array([ 5. , 13. , 1. , 1.41421356])
For the same vectors, compute the p-norm for p = [1, 2, inf]:
>>> vnorm([[3, 4], [5, 12], [0, 1], [1, 1]], [[1], [2], [np.inf]])
array([[ 7. , 17. , 1. , 2. ],
[ 5. , 13. , 1. , 1.41421356],
[ 4. , 12. , 1. , 1. ]])
vnorm handles complex numbers. Here we compute the norm of z
with orders 1, 2, 3, and inf. (Note that abs(z) is [2, 5, 0, 14].)
>>> z = np.array([-2j, 3+4j, 0, 14])
>>> vnorm(z, [1, 2, 3, np.inf])
array([21. , 15. , 14.22263137, 14. ])
vdot(x, y) is the vector dot product of the real floating point vectors
x and y. It is a gufunc with signature (n),(n)->().
>>> import numpy as np
>>> from ufunclab import vdot
>>> x = np.array([[1, -2, 3],
... [4, 5, 6]])
>>> y = np.array([[-1, 0, 3],
[1, 1, 1]])
>>> vdot(x, y) # Default axis is -1.
array([ 8., 15.])
>>> vdot(x, y, axis=0)
array([ 3., 5., 15.])
pearson_corr(x, y) computes Pearson's product-moment correlation coefficient.
It is a gufunc with shape signature (n),(n)->().
>>> import numpy as np
>>> from ufunclab import pearson_corr
>>> x = np.array([1.0, 2.0, 3.5, 7.0, 8.5, 10.0, 11.0])
>>> y = np.array([10, 11.5, 11.4, 13.6, 15.1, 16.7, 15.0])
>>> pearson_corr(x, y)
0.9506381287828245
In the following example, a trivial dimension is added to the array a before
passing it to pearson_corr, so the inputs are compatible for broadcasting.
The correlation coefficient of each row of a with each row of b is computed,
giving a result with shape (3, 2).
>>> a = np.array([[2, 3, 1, 3, 5, 8, 8, 9],
... [3, 3, 1, 2, 2, 4, 4, 5],
... [2, 5, 1, 2, 2, 3, 3, 8]])
>>> b = np.array([[9, 8, 8, 7, 4, 4, 1, 2],
... [8, 9, 9, 6, 5, 7, 3, 4]])
>>> pearson_corr(np.expand_dims(a, 1), b)
array([[-0.92758645, -0.76815464],
[-0.65015428, -0.53015896],
[-0.43575108, -0.32925148]])
wjaccard is a gufunc with shape signature (n),(n)->() that computes
the weighted Jaccard index, which is defined to be
Jw(x, y) = min(x, y).sum() / max(x, y).sum()
>>> import numpy as np
>>> from ufunclab import wjaccard
>>> x = np.array([0.9, 1.0, 0.7, 0.0, 0.8, 0.6])
>>> y = np.array([0.3, 1.0, 0.9, 0.6, 1.0, 0.2])
>>> wjaccard(x, y)
0.6
cross2(u, v) is a gufunc with signature (2),(2)->(). It computes
the 2-d cross product that returns a scalar. That is, cross2([u0, u1], [v0, v1])
is u0*v1 - u1*v0. The calculation is the same as that of numpy.cross,
but cross2 is restricted to 2-d inputs.
For example,
>>> import numpy as np
>>> from ufunclab import cross2
>>> cross2([1, 2], [5, 3])
-7
>>> cross2([[1, 2], [6, 0]], [[5, 3], [2, 3]])
array([-7, 18])
>>> cross2([1j, 3], [-1j, 2+3j])
(-3+5j)
In the following, a and b are object arrays; a has shape (2,),
and b has shape (3, 2). The result of cross2(a, b) has shape
(3,).
>>> from fractions import Fraction as F
>>> a = np.array([F(1, 3), F(2, 7)])
>>> b = np.array([[F(7, 4), F(6, 7)], [F(2, 5), F(-3, 7)], [1, F(1, 4)]])
>>> cross2(a, b)
array([Fraction(-3, 14), Fraction(-9, 35), Fraction(-17, 84)],
dtype=object)
cross3(u, v) is a gufunc with signature (3),(3)->(3). It computes
the 3-d vector cross product (like numpy.cross, but specialized to the
case of 3-d vectors only).
For example,
>>> import numpy as np
>>> from ufunclab import cross3
>>> u = np.array([1, 2, 3])
>>> v = np.array([2, 2, -1])
>>> cross3(u, v)
array([-8, 7, -2])
In the following, x has shape (5, 3), and y has shape (2, 1, 3).
The result of cross3(x, y) has shape (2, 5, 3).
>>> x = np.arange(15).reshape(5, 3)
>>> y = np.round(10*np.sin(np.linspace(0, 2, 6))).reshape(2, 1, 3)
>>> x
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11],
[12, 13, 14]])
>>> y
array([[[ 0., 4., 7.]],
[[ 9., 10., 9.]]])
>>> cross3(x, y)
array([[[ -1., 0., 0.],
[ 8., -21., 12.],
[ 17., -42., 24.],
[ 26., -63., 36.],
[ 35., -84., 48.]],
[[-11., 18., -9.],
[-14., 18., -6.],
[-17., 18., -3.],
[-20., 18., 0.],
[-23., 18., 3.]]])
tri_area(p) is a gufunc with signature (3, n) - > (). It computes the
area of a triangle defined by three points in n-dimensional space.
>>> import numpy as np
>>> from ufunclab import tri_area
`p` has shape (2, 3, 4). It contains the vertices
of two triangles in 4-dimensional space.
>>> p = np.array([[[0.0, 0.0, 0.0, 6.0],
[1.0, 2.0, 3.0, 6.0],
[0.0, 2.0, 2.0, 6.0]],
[[1.5, 1.0, 2.5, 2.0],
[4.0, 1.0, 0.0, 2.5],
[2.0, 1.0, 2.0, 2.5]]])
>>> tri_area(p)
array([1.73205081, 0.70710678])
fillnan1d(x) is a gufunc with signature (i)->(i). It uses linear
interpolation to replace occurrences of nan in x.
>>> import numpy as np
>>> from ufunclab import fillnan1d
>>> x = np.array([1.0, 2.0, np.nan, np.nan, 3.5, 5.0, np.nan, 7.5])
>>> fillnan1d(x)
array([1. , 2. , 2.5 , 3. , 3.5 , 5. , 6.25, 7.5 ])
nan values at the ends of x are replaced with the nearest non-nan
value:
>>> x = np.array([np.nan, 2.0, np.nan, 5.0, np.nan, np.nan])
>>> fillnan1d(x)
array([2. , 2. , 3.5, 5. , 5. , 5. ])
This plot of the result of applying fillnan1d(x) to a bigger sample is
generated by the script examples/fillnan1d_demo.py:
linear_interp1d(x, xp, fp) is a gufunc with signature (),(n),(n)->(). It
is like numpy.interp, but as a gufunc, it will broadcast its arguments.
Currently the function provides just the basic interpolation function. It does
not extrapolate, so any x values outside of the interval [xp[0], xp[-1]] will
result in nan. The period option of numpy.interp is also not implemented.
The function has a special code path in which the computation of the indices
where the values in x lie in xp is done once for each input. To activate
this code path, the appropriate shapes of the input arrays must be used. Some
familiarity with the broadcasting behavior of gufuncs is necessary to make the
most of this fast code path. The arrays must be shaped so that in the internal
gufunc loop, the strides associated with the x and xp arrays are 0.
Some examples of the use of linear_interp1d follow.
>>> import numpy as np
>>> from ufunclab import linear_interp1d
This example is the same as numpy.interp. xp and fp are the known values,
and we want to evaluate the interpolated function at x = [0, 0.25, 1, 2, 5, 6].
>>> xp = np.array([0, 1, 3, 5, 8])
>>> fp = np.array([10, 15, 15, 20, 80])
>>> x = np.array([0, 0.25, 1, 2, 5, 6.5])
>>> linear_interp1d(x, xp, fp)
array([10. , 11.25, 15. , 15. , 20. , 50. ])
>>> np.interp(x, xp, fp)
array([10. , 11.25, 15. , 15. , 20. , 50. ])
With broadcasting, we can interpolate multiple functions (i.e. multiple 1-d arrays) in one call.
For example, in the array fp3 below, we want to treat each column as a
separate function to be interpolated (note that the first column is the
same as fp above):
>>> fp3 = np.array([[10, 15, 10],
... [15, 14, 20],
... [15, 13, 40],
... [20, 12, 10],
... [80, 11, 0]])
We can do this with linear_interp1d, but we have to adjust the shapes
of the inputs to broadcast correctly:
>>> linear_interp1d(x[:, None], xp, fp3.T)
array([[10. , 15. , 10. ],
[11.25, 14.75, 12.5 ],
[15. , 14. , 20. ],
[15. , 13.5 , 30. ],
[20. , 12. , 10. ],
[50. , 11.5 , 5. ]])
The script linear_interp1d_demo.py in the examples directory
provides a demonstration of the use of linear_interp1d to interpolate
a parametric curve in two dimensions using arclength as the parameter.
It generates the following plot:
backlash(x, deadband, initial), a gufunc with signature (i),(),()->(i),
computes the "backlash" response of a signal; see the Wikipedia article
Backlash (engineering).
The function emulates the
backlash block
of Matlab's Simulink library.
For example,
>>> import numpy as np
>>> from ufunclab import backlash
>>> x = np.array([0, 0.5, 1, 1.1, 1.0, 1.5, 1.4, 1.2, 0.5])
>>> deadband = 0.4
>>> initial = 0
>>> backlash(x, deadband, initial)
array([0. , 0.3, 0.8, 0.9, 0.9, 1.3, 1.3, 1.3, 0.7])
The script backlash_demo.py in the examples directory generates
the plot
hysteresis_relay(x, low_threshold, high_threshold, low_value, high_value, init),
a gufunc with signature (i),(),(),(),(),()->(i), passes x through a relay with
hysteresis (like a Schmitt trigger).
The function is similar to the
relay block
of Matlab's Simulink library.
The script hysteresis_relay_demo.py in the examples directory generates
the plot
sosfilter(sos, x) is a gufunc with signature (m,6),(n)->(n).
The function applies a discrete time linear filter to the input
array x. The array sos with shape (m,6) represents the
linear filter using the second order sections format.
The function is like scipy.signal.sosfilt, but this version does
not accept the zi parameter. See sosfilter_ic for a function
that accepts zi.
The script sosfilter_demo.py in the examples directory generates
the plot
sosfilter_ic(sos, x, zi) is a gufunc with signature
(m,6),(n),(m,2)->(n),(m,2). Like sosfilter, the function applies
a discrete time linear filter to the input array x. The array sos
with shape (m,6) represents the linear filter using the second order
sections format.
This function is like scipy.signal.sosfilt, but for sosfilter_ic,
the zi parameter is required. Also, because sosfilter_ic is a gufunc,
it uses the gufunc rules for broadcasting. scipy.signal.sosfilt handles
broadcasting of the zi parameter differently.
sosfilter_ic_contig(sos, x, zi) is a gufunc with signature
(m,6),(n),(m,2)->(n),(m,2). This function has the same inputs and
performs the same calculation as sosfilter_ic, but it assumes that the
array inputs are all C-contiguous. It does not verify this; if an array
input is not C-contiguous, the results will be incorrect, and the program
might crash.
multivariate_logbeta(x) is a gufunc with signature (n)->() that computes
the logarithm of the multivariate beta function.
>>> import numpy as np
>>> from ufunclab import multivariate_logbeta
>>> x = np.array([1, 2.5, 7.25, 3])
>>> multivariate_logbeta(x)
-13.87374699005739
Compare to
>>> from scipy.special import gammaln
>>> gammaln(x).sum() - gammaln(x.sum())
-13.87374699005739
gendot creates a new gufunc (with signature (i),(i)->()) that is
the composition of two ufuncs. The first ufunc must be an element-wise
ufunc with two inputs and one output. The second must be either another
element-wise ufunc with two inputs and one output, or a gufunc with
signature (i)->().
The name gendot is from "generalized dot product". The standard
dot product is the composition of element-wise multiplication and
reduction with addition. The prodfunc and sumfunc arguments of
gendot take the place of multiplication and addition.
For example, to take the element-wise minimum of two 1-d arrays, and then take the maximum of the result:
>>> import numpy as np
>>> from ufunclab import gendot
>>> minmaxdot = gendot(np.minimum, np.maximum)
>>> a = np.array([1.0, 2.5, 0.3, 1.9, 3.0, 1.8])
>>> b = np.array([0.5, 1.1, 0.9, 2.1, 0.3, 3.0])
>>> minmaxdot(a, b)
1.9
minmaxdot is a gufunc with signature (i),(i)->(); the type
signatures of the gufunc loop functions were derived by matching
the signatures of the ufunc loop functions for np.minimum and
np.maximum:
>>> minmaxdot.signature
'(i),(i)->()'
>>> print(minmaxdot.types)
['??->?', 'bb->b', 'BB->B', 'hh->h', 'HH->H', 'ii->i', 'II->I', 'll->l',
'LL->L', 'qq->q', 'QQ->Q', 'ee->e', 'ff->f', 'dd->d', 'gg->g', 'FF->F',
'DD->D', 'GG->G', 'mm->m', 'MM->M']
gendot is experimental, and might not be useful in many applications.
We could do the same calculation as minmaxdot with, for example,
np.maximum.reduce(np.minimum(a, b)), and in fact, the pure NumPy
version is faster than minmaxdot(a, b) for large (and even moderately
sized) 1-d arrays. An advantage of the gendot gufunc is that it does
not create an intermediate array when broadcasting takes place. For
example, with inputs x and y with shapes (20, 10000000) and
(10, 1, 10000000), the equivalent of minmaxdot(x, y) can be computed
with np.maximum.reduce(np.minimum(x, y), axis=-1), but np.minimum(x, y)
creates an array with shape (10, 20, 10000000). Computing the result
with minmaxdot(x, y) does not create the temporary intermediate array.
ufunc_inspector(func) prints information about a NumPy ufunc.
For example,
>>> import numpy as np
>>> from ufunclab import ufunc_inspector
>>> np.__version__
'1.24.0'
>>> ufunc_inspector(np.hypot)
'hypot' is a ufunc.
nin = 2, nout = 1
ntypes = 5
loop types:
0: ( 23, 23) -> 23 (ee->e) PyUFunc_ee_e_As_ff_f
1: ( 11, 11) -> 11 (ff->f) PyUFunc_ff_f
2: ( 12, 12) -> 12 (dd->d) PyUFunc_dd_d
3: ( 13, 13) -> 13 (gg->g) PyUFunc_gg_g
4: ( 17, 17) -> 17 (OO->O) PyUFunc_OO_O_method
(The output will likely change as the code develops.)
>>> ufunc_inspector(np.sqrt)
'sqrt' is a ufunc.
nin = 1, nout = 1
ntypes = 10
loop types:
0: 23 -> 23 (e->e) PyUFunc_e_e_As_f_f
1: 11 -> 11 (f->f) not generic (or not in the checked generics)
2: 12 -> 12 (d->d) not generic (or not in the checked generics)
3: 11 -> 11 (f->f) PyUFunc_f_f
4: 12 -> 12 (d->d) PyUFunc_d_d
5: 13 -> 13 (g->g) PyUFunc_g_g
6: 14 -> 14 (F->F) PyUFunc_F_F
7: 15 -> 15 (D->D) PyUFunc_D_D
8: 16 -> 16 (G->G) PyUFunc_G_G
9: 17 -> 17 (O->O) PyUFunc_O_O_method
>>> ufunc_inspector(np.add)
'add' is a ufunc.
nin = 2, nout = 1
ntypes = 22
loop types:
0: ( 0, 0) -> 0 (??->?) not generic (or not in the checked generics)
1: ( 1, 1) -> 1 (bb->b) not generic (or not in the checked generics)
2: ( 2, 2) -> 2 (BB->B) not generic (or not in the checked generics)
3: ( 3, 3) -> 3 (hh->h) not generic (or not in the checked generics)
4: ( 4, 4) -> 4 (HH->H) not generic (or not in the checked generics)
5: ( 5, 5) -> 5 (ii->i) not generic (or not in the checked generics)
6: ( 6, 6) -> 6 (II->I) not generic (or not in the checked generics)
7: ( 7, 7) -> 7 (ll->l) not generic (or not in the checked generics)
8: ( 8, 8) -> 8 (LL->L) not generic (or not in the checked generics)
9: ( 9, 9) -> 9 (qq->q) not generic (or not in the checked generics)
10: ( 10, 10) -> 10 (QQ->Q) not generic (or not in the checked generics)
11: ( 23, 23) -> 23 (ee->e) not generic (or not in the checked generics)
12: ( 11, 11) -> 11 (ff->f) not generic (or not in the checked generics)
13: ( 12, 12) -> 12 (dd->d) not generic (or not in the checked generics)
14: ( 13, 13) -> 13 (gg->g) not generic (or not in the checked generics)
15: ( 14, 14) -> 14 (FF->F) not generic (or not in the checked generics)
16: ( 15, 15) -> 15 (DD->D) not generic (or not in the checked generics)
17: ( 16, 16) -> 16 (GG->G) not generic (or not in the checked generics)
18: ( 21, 22) -> 21 (Mm->M) not generic (or not in the checked generics)
19: ( 22, 22) -> 22 (mm->m) not generic (or not in the checked generics)
20: ( 22, 21) -> 21 (mM->M) not generic (or not in the checked generics)
21: ( 17, 17) -> 17 (OO->O) PyUFunc_OO_O
Here's a collection of resources for learning about the C API for ufuncs.
- Universal functions (ufunc)
- UFunc API
- Generalized Universal Function API
- NEP 5 — Generalized Universal Functions
- NEP 20 — Expansion of Generalized Universal Function Signatures
- Universal functions,
part of the NumPy C Code Explanations
- In particular, the section "Function call" explains when the GIL is released.
- When implementing inner loops for many NumPy dtypes, the
NumPy distutils
template preprocessor
is a useful tool. (See the "Other files"
section for the syntax that would be used in, say, a C file.)
However, the
distutilssubpackage of NumPy is deprecated. - Some relevant NumPy source code, if you want to dive deep:
PyUFuncObjectalong with related C types and macros are defined innumpy/numpy/core/include/numpy/ufuncobject.h.PyUFunc_FromFuncAndDataandPyUFunc_FromFuncAndDataAndSignatureAndIdentityare defined in the filenumpy/numpy/core/src/umath/ufunc_object.c.
- Section of the SciPy Lecture Notes on ufuncs:
- Data Type API -- a handy reference.