floating point errors

[todbot]

CircuitPython version

Adafruit CircuitPython 9.1.0-beta.1-18-g781c577745 on 2024-05-05; Adafruit QT Py RP2040 with rp2040

Code/REPL

print("Hello, Pi Pico!")

print("0.004 = ", 0.004)
print("1000 = ", 1000)
print("0.004 * 1000 = ", 0.004 * 1000)
print("int(0.004 * 1000) = ", int(0.004 * 1000))
print('"%f" % (0.004 * 1000) = ', "%f" % (0.004 * 1000))

Behavior

Prints out a "3" instead of a "4" on the second-to-last line and "3.999998" instead of "4.000000"

Hello, Pi Pico!
0.004 =  0.004
1000 =  1000
0.004 * 1000 =  4.0
int(0.004 * 1000) =  3
"%f" % (0.004 * 1000) =  3.999998

Description

The crux of the issue is that int(0.004 * 1000) prints out "3" instead of "4".

It's unclear to me why it does this, both numbers have pretty simple representations in IEEE754 (using https://www.h-schmidt.net/FloatConverter/IEEE754.html, 0.004 is actually 0.0040000001899898052215576171875, 1000 is actually 1000)

This error does not exist in MicroPython v1.22.0 on 2023-12-27.

Here are two wokwis to compare:

• CircuitPython 8.0.2 -- https://wokwi.com/projects/397539939082977281
• Micropython v1.22.0 -- https://wokwi.com/projects/397539826736442369

Additional information

No response

[justmobilize]

@todbot I do see if I remove that last bit of mantissa (which CP doesn't have) it's 0.0039999997, which would match your results...

[dhalbert]

Remove that last two bits for what CircuitPython has. As I mentioned in discord, the CircuitPython default for objects is OBJ_REPR_C. MicroPython defaults to OBJ_REPR_A. See

circuitpython/py/mpconfig.h

Lines 105 to 107 in d8ca842

	/* Object representation */
	// A MicroPython object is a machine word having the following form:

and following.

See also #5802, #1220, #566, #230, and the MicroPython issues those issues reference.

This is always going to be a problem with floats. It's worse the fewer bits of precision you have. MicroPython has improved its printing of floats over the years, and of course we've incorporated that, but there is is a trade-off between "reasonable" and seemingly more exact.

We could consider going to boxed floats, but for lots of floats in lists (not array.array), it uses up a lot of storage.