int(1.0/7*7) == 0 ???

[notro]

I'm working on converting the datetime module and came across this.
Is it a known limitation or a bug?

Adafruit CircuitPython 3.0.1 on 2018-08-21; Adafruit CircuitPlayground Express with samd21g18
>>> int(1.0)
1
>>> 1.0/7*7
1.0
>>> int(1.0/7*7)
0
>>>

Adafruit CircuitPython patchbase-7-g216f0f952-dirty on 2018-09-30; Adafruit Metro M4 Express with samd51j19
>>> int(1.0)
1
>>> 1.0/7*7
1.0
>>> int(1.0/7*7)
0
>>>

Python 3.4.2 (default, Oct 19 2014, 13:31:11)
[GCC 4.9.1] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> int(1.0/7*7)
1
>>>

[jerryneedell]

FYI

Press any key to enter the REPL. Use CTRL-D to reload.
Adafruit CircuitPython 4.0.0-alpha.1-5-g48a3aafdd on 2018-09-25; PCA10059 nRF52840 Dongle with nRF52840
>>> 1.0/7*7
1.0
>>> int(1.0/7*7)
1
>>> 

but

Adafruit CircuitPython 4.0.0-alpha.1-2-gf2c960e89 on 2018-09-21; Adafruit Metro M4 Express with samd51j19
>>> 1.0/7*7
1.0
>>> int(1.0/7*7)
0
>>> 

[dhalbert]

It may print as 1.0 but it's not quite 1.0:

Adafruit CircuitPython 3.0.2 on 2018-09-14; Adafruit Feather M4 Express with samd51j19
>>> int(1.0/7*7)
0
>>> 1.0/7*7
1.0
>>> 1.0 - (1.0/7*7)
2.38419e-07

[jerryneedell]

interesting

Auto-reload is on. Simply save files over USB to run them or enter REPL to disable.

Press any key to enter the REPL. Use CTRL-D to reload.
Adafruit CircuitPython 4.0.0-alpha.1-5-g48a3aafdd on 2018-09-25; PCA10059 nRF52840 Dongle with nRF52840
>>> 1.0 - (1.0/7*7)
0.0

[dhalbert]

4.0.0 on M4 is still the same:

Adafruit CircuitPython 4.0.0-alpha.1-4-gf09537bbc on 2018-09-30; Adafruit Feather M4 Express with samd51j19
>>> int(1.0/7*7)
0
>>> 1.0 - (1.0/7*7)
2.38419e-07

Not sure what the difference is: it might be a compilation flag or some setup of the floating point unit. I thought both nRF and Atmel used the same conversion routines, so I'm a little surprised, but ma7be the don't.

[dhalbert]

I thought this might be because we are using the MicroPython-supplied libm on atmel-samd. But if I use the toolchain libm instead (which consumes an extra ~5600 bytes), I still get the same results on the Feather M4.

[notro]

This is the gist of the code from datetime.py that I work on:

import math as _math
class timedelta:

    def __new__(cls, days=0, seconds=0, microseconds=0,
                milliseconds=0, minutes=0, hours=0, weeks=0):
        days += weeks*7

        if isinstance(days, float):
            dayfrac, days = _math.modf(days)
            d = int(days)
        else:
            d = days

        self = object.__new__(cls)
        self._days = d
        return self

The test wants these to be equal:

>>> timedelta(weeks=1.0/7)._days
0
>>> timedelta(days=1)._days
1

Any idea how I can achieve that in a general way under the current circumstances?

[jepler]

In floating point math, it's not generally true that int(1 / i * i) == 1 for all integers i. On modern 64-bit Linux, the first positive number which is an exception is 49. On CircuitPython, since floating point numbers are stored in a more limited precision (slightly less precision than a 32-bit "float" on modern desktop computers, and a LOT less precision than the standard 64-bit "double" precision float of desktop Python), this problem affects smaller numbers such as 7.

Here's a similar "problem" on desktop Python 3.5.3:

>>> import datetime
>>> datetime.timedelta(weeks=1/11)*11
datetime.timedelta(7, 0, 2)

In this case, the duration is reported as 2 microseconds longer than 7 days.

If this problem is arising from a doctest, I suggest that perhaps the doctest could be rewritten so that it asserts that a timedelta of 1/7 weeks is close to a timedelta of 1 day, instead of asserting that it is equal exactly. For instance, you could assert that both are within 1 second of each other via the total_seconds() method, which should be true on python and circuitpython platforms.

[notro]

Thanks for the explanation, floats and its lack of precision keep tripping me up.
I'll see what I can do with the tests.

Thanks all.

Example of the (im)precision showing the value I need to use to get exactly one day as a result:

>>> datetime.timedelta(weeks=1.0000005/7)
datetime.timedelta(1)

[notro]

total_seconds() was useful for the test case:

 class TestTimeDelta(HarmlessMixedComparison, unittest.TestCase):

     def test_constructor(self):

         # Check float args to constructor
+        def aeq(td1, td2):  ### Handle float imprecision
+            self.assertAlmostEqual(td1.total_seconds(), td2.total_seconds(), delta=1)
-        eq(td(weeks=1.0/7), td(days=1))
+        aeq(td(weeks=1.0/7), td(days=1))
-        eq(td(days=1.0/24), td(hours=1))
+        aeq(td(days=1.0/24), td(hours=1))
-        eq(td(hours=1.0/60), td(minutes=1))
+        aeq(td(hours=1.0/60), td(minutes=1))
         eq(td(minutes=1.0/60), td(seconds=1))
         eq(td(seconds=0.001), td(milliseconds=1))
         eq(td(milliseconds=0.001), td(microseconds=1))