[BUG] Negating a boolean Series changes its dtype to int8

[shwina]

In Pandas, negating a bool Series behaves as follows:

>>> -pd.Series([True, False])
0    False
1     True
dtype: bool

Whereas in cuDF:

>>> -cudf.Series([True, False]) 
0   -1
1    0
dtype: int8

[shwina]

Should we match Pandas behaviour here?

[wence-]

Pandas doesn't promote types in uops, so unsigned types stay unsigned, whereas cudf does:

import pandas as pd
import numpy as np
for dtype in ["bool", "uint8", "int8", "uint64", "int64"]:
    if dtype == "bool":
        dtype_max = True
        dtype_min = False
    else:
        dtype = np.dtype(dtype)
        dtype_max = np.iinfo(dtype).max
        dtype_min = np.iinfo(dtype).min
    s = pd.Series([1], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = pd.Series([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = pd.Series([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)
# bool bool
# bool bool
# bool bool
# uint8 uint8
# uint8 uint8
# uint8 uint8
# int8 int8
# int8 int8
# int8 int8
# uint64 uint64
# uint64 uint64
# uint64 uint64
# int64 int64
# int64 int64
# int64 int64

import cudf
import numpy as np
for dtype in ["bool", "uint8", "int8", "uint64", "int64"]:
    if dtype == "bool":
        dtype_max = True
        dtype_min = False
    else:
        dtype = np.dtype(dtype)
        dtype_max = np.iinfo(dtype).max
        dtype_min = np.iinfo(dtype).min
    s = cudf.Series([1], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = cudf.Series([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = cudf.Series([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)

# bool int8
# bool int8
# bool int8
# uint8 int16
# uint8 int16
# uint8 int16
# int8 int8
# int8 int8
# int8 int8
# uint64 float64
# uint64 float64
# uint64 float64
# int64 int64
# int64 int64
# int64 int64

As we can see, cudf promotes to the next signed type that is wide enough to hold the range of values that the unsigned type has (including promoting uint64 to float64, and losing precision).

Finally, numpy raises TypeError on unary minus of booleans, suggesting one use ~ instead, and otherwise does not promote types

import numpy as np
for dtype in ["uint8", "int8", "uint64", "int64"]:
    dtype = np.dtype(dtype)
    dtype_max = np.iinfo(dtype).max
    dtype_min = np.iinfo(dtype).min
    s = np.asarray([1], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = np.asarray([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)
    s = np.asarray([dtype_max], dtype=dtype)
    print(s.dtype, (-s).dtype)

# uint8 uint8
# uint8 uint8
# uint8 uint8
# int8 int8
# int8 int8
# int8 int8
# uint64 uint64
# uint64 uint64
# uint64 uint64
# int64 int64
# int64 int64
# int64 int64

I think we should match pandas here.

[wence-]

Note in general that type promotion involving booleans probably needs special case handling because I am not sure we do the right thing for binops now either.

[vyasr]

It looks like we resolved the original issue at some point here:

In [80]: -cudf.Series([True, False])
Out[80]:
0    False
1     True
dtype: bool