log2, log10, acosh and tgamma don't return NaN when they should #1256

New issue

Open

opened 2023-02-13 23:37:29 -06:00 by dpgeorge · 4 comments

dpgeorge commented

2023-02-13 23:37:29 -06:00

(Migrated from github.com)

The following functions evaluated at the given points return a finite or infinite result when then should return Nan:

log2(-INFINITY) returns -INFINITY but should return NaN
log10(-INFINITY) returns -INFINITY but should return NaN
acosh(-1.0) returns 0.0 but should return NaN (it should return NaN for all arguments less than 1)
acosh(-INFINITY) returns INFINITY but should return NaN
tgamma(-INFINITY) returns INFINITY but should return NaN

I'm not sure exactly what the policy is for domain errors, but when PICO_FLOAT_PROPAGATE_NANS is enabled a lot of other cases of NaN are handled, so one would expect the above to also be handled correctly.

Edit: possibly also the following:

sqrt(-INFINITY) returns -INFINITY but should return NaN
log(-INFINITY) returns -INFINITY but should return NaN

The following functions evaluated at the given points return a finite or infinite result when then should return Nan: - `log2(-INFINITY)` returns `-INFINITY` but should return NaN - `log10(-INFINITY)` returns `-INFINITY` but should return NaN - `acosh(-1.0)` returns `0.0` but should return NaN (it should return NaN for all arguments less than 1) - `acosh(-INFINITY)` returns `INFINITY` but should return NaN - `tgamma(-INFINITY)` returns `INFINITY` but should return NaN I'm not sure exactly what the policy is for domain errors, but when `PICO_FLOAT_PROPAGATE_NANS` is enabled a lot of other cases of NaN are handled, so one would expect the above to also be handled correctly. Edit: possibly also the following: - `sqrt(-INFINITY)` returns `-INFINITY` but should return NaN - `log(-INFINITY)` returns `-INFINITY` but should return NaN

kilograham commented

2023-02-15 15:03:33 -06:00

(Migrated from github.com)

yes, there seems to be a mixture - initially the policy was intended to be PROPAGATE_NANs should take NAN input to NAN output, but we clearly have code in some places to fix this.

I'm surprised log(-INFINITY) isn't listed...

can you point me at the test for this in MP

yes, there seems to be a mixture - initially the policy was intended to be PROPAGATE_NANs should take NAN input to NAN output, but we clearly have code in some places to fix this. I'm surprised log(-INFINITY) isn't listed... can you point me at the test for this in MP

dpgeorge commented

2023-02-15 16:55:36 -06:00

(Migrated from github.com)

tests/float/math_domain_special.py is the test.

That tests also see some strange behaviour of math.asinh(0) returning -0.0 when it should be 0.0, but I couldn't track that bug down because when I ran asinhf(0.0) in C code it returned 0.0 as expected.

`tests/float/math_domain_special.py` is the test. That tests also see some strange behaviour of `math.asinh(0)` returning `-0.0` when it should be `0.0`, but I couldn't track that bug down because when I ran `asinhf(0.0)` in C code it returned `0.0` as expected.

dpgeorge commented

2023-02-15 17:14:18 -06:00

(Migrated from github.com)

I'm surprised log(-INFINITY) isn't listed...

Surprisingly we don't have a test for this... I'm in the process of adding one but it seems that it does pass on rp2 because we handle math.log(x[, base]) differently to log2 and log10: log explicitly checks that it's argument is positive.

> I'm surprised log(-INFINITY) isn't listed... Surprisingly we don't have a test for this... I'm in the process of adding one but it seems that it does pass on rp2 because we handle `math.log(x[, base])` differently to log2 and log10: log explicitly checks that it's argument is positive.

dpgeorge commented

2023-02-15 17:23:53 -06:00

(Migrated from github.com)

To run the tests on MP:

$ cd tests
$ ./run-tests.py --target rp2 -d float

Then to see the diff of the expected (left) vs actual (right) results:

$ diff results/float_cmath_fun.py.exp results/float_cmath_fun.py.out 
113c113
< complex(-109.71, 55.602)
---
> complex(-109.72, 55.601)

$ diff results/float_math_domain.py.exp results/float_math_domain.py.out 
16c16
< sqrt(-inf) = ValueError
---
> sqrt(-inf) = -inf

$ diff results/float_math_domain_special.py.exp results/float_math_domain_special.py.out
7c7
< log2(-inf) = ValueError
---
> log2(-inf) = -inf
12c12
< log10(-inf) = ValueError
---
> log10(-inf) = -inf
26c26
< acosh(-1.0000) = ValueError
---
> acosh(-1.0000) = 0.0000
30c30
< acosh(-inf) = ValueError
---
> acosh(-inf) = inf
48c48
< gamma(-inf) = ValueError
---
> gamma(-inf) = inf

$ diff results/float_math_fun_special.py.exp results/float_math_fun_special.py.out
60c60
< asinh(0) = 0
---
> asinh(0) = -0

(The first one there is #1255.)

To run the tests on MP: ```sh $ cd tests $ ./run-tests.py --target rp2 -d float ``` Then to see the diff of the expected (left) vs actual (right) results: ```sh $ diff results/float_cmath_fun.py.exp results/float_cmath_fun.py.out 113c113 < complex(-109.71, 55.602) --- > complex(-109.72, 55.601) $ diff results/float_math_domain.py.exp results/float_math_domain.py.out 16c16 < sqrt(-inf) = ValueError --- > sqrt(-inf) = -inf $ diff results/float_math_domain_special.py.exp results/float_math_domain_special.py.out 7c7 < log2(-inf) = ValueError --- > log2(-inf) = -inf 12c12 < log10(-inf) = ValueError --- > log10(-inf) = -inf 26c26 < acosh(-1.0000) = ValueError --- > acosh(-1.0000) = 0.0000 30c30 < acosh(-inf) = ValueError --- > acosh(-inf) = inf 48c48 < gamma(-inf) = ValueError --- > gamma(-inf) = inf $ diff results/float_math_fun_special.py.exp results/float_math_fun_special.py.out 60c60 < asinh(0) = 0 --- > asinh(0) = -0 ``` (The first one there is #1255.)