Skip to content

MAINT: simplify torch dtype promotion #303

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Apr 15, 2025
Merged

MAINT: simplify torch dtype promotion #303

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
845d8b5
ENH: torch dtype promotions
crusaderky Apr 3, 2025
39d285c
tweak _result_type
crusaderky Apr 10, 2025
1ae5c1b
Revert uint promotions
crusaderky Apr 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
99 changes: 40 additions & 59 deletions array_api_compat/torch/_aliases.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,54 +35,33 @@
torch.complex128,
}

_promotion_table = {
# bool
(torch.bool, torch.bool): torch.bool,
_promotion_table = {
# ints
(torch.int8, torch.int8): torch.int8,
(torch.int8, torch.int16): torch.int16,
(torch.int8, torch.int32): torch.int32,
(torch.int8, torch.int64): torch.int64,
(torch.int16, torch.int8): torch.int16,
(torch.int16, torch.int16): torch.int16,
(torch.int16, torch.int32): torch.int32,
(torch.int16, torch.int64): torch.int64,
(torch.int32, torch.int8): torch.int32,
(torch.int32, torch.int16): torch.int32,
(torch.int32, torch.int32): torch.int32,
(torch.int32, torch.int64): torch.int64,
(torch.int64, torch.int8): torch.int64,
(torch.int64, torch.int16): torch.int64,
(torch.int64, torch.int32): torch.int64,
(torch.int64, torch.int64): torch.int64,
# uints
(torch.uint8, torch.uint8): torch.uint8,
# ints and uints (mixed sign)
(torch.int8, torch.uint8): torch.int16,
(torch.int16, torch.uint8): torch.int16,
(torch.int32, torch.uint8): torch.int32,
(torch.int64, torch.uint8): torch.int64,
(torch.uint8, torch.int8): torch.int16,
(torch.uint8, torch.int16): torch.int16,
(torch.uint8, torch.int32): torch.int32,
(torch.uint8, torch.int64): torch.int64,
# floats
(torch.float32, torch.float32): torch.float32,
(torch.float32, torch.float64): torch.float64,
(torch.float64, torch.float32): torch.float64,
(torch.float64, torch.float64): torch.float64,
# complexes
(torch.complex64, torch.complex64): torch.complex64,
(torch.complex64, torch.complex128): torch.complex128,
(torch.complex128, torch.complex64): torch.complex128,
(torch.complex128, torch.complex128): torch.complex128,
# Mixed float and complex
(torch.float32, torch.complex64): torch.complex64,
(torch.float32, torch.complex128): torch.complex128,
(torch.float64, torch.complex64): torch.complex128,
(torch.float64, torch.complex128): torch.complex128,
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(complex, float) use cases were missing

}

_promotion_table.update({(b, a): c for (a, b), c in _promotion_table.items()})
_promotion_table.update({(a, a): a for a in _array_api_dtypes})


def _two_arg(f):
@_wraps(f)
Expand Down Expand Up @@ -150,13 +129,18 @@ def result_type(
return _reduce(_result_type, others + scalars)


def _result_type(x, y):
def _result_type(
x: Array | DType | bool | int | float | complex,
y: Array | DType | bool | int | float | complex,
) -> DType:
if not (isinstance(x, _py_scalars) or isinstance(y, _py_scalars)):
xdt = x.dtype if not isinstance(x, torch.dtype) else x
ydt = y.dtype if not isinstance(y, torch.dtype) else y
xdt = x if isinstance(x, torch.dtype) else x.dtype
ydt = y if isinstance(y, torch.dtype) else y.dtype

if (xdt, ydt) in _promotion_table:
try:
return _promotion_table[xdt, ydt]
except KeyError:
pass

# This doesn't result_type(dtype, dtype) for non-array API dtypes
# because torch.result_type only accepts tensors. This does however, allow
Expand Down Expand Up @@ -301,27 +285,35 @@ def _reduce_multiple_axes(f, x, axis, keepdims=False, **kwargs):
out = torch.unsqueeze(out, a)
return out


def _sum_prod_no_axis(x: Array, dtype: DType | None) -> Array:
"""
Implements `sum(..., axis=())` and `prod(..., axis=())`.

Works around https://github.com/pytorch/pytorch/issues/29137
"""
if dtype is not None:
return x.clone() if dtype == x.dtype else x.to(dtype)

# We can't upcast uint8 according to the spec because there is no
# torch.uint64, so at least upcast to int64 which is what prod does
# when axis=None.
if x.dtype in (torch.uint8, torch.int8, torch.int16, torch.int32):
return x.to(torch.int64)

return x.clone()
Copy link
Member

@ev-br ev-br Apr 15, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note to self: this looks scary, but is in fact just a refactoring. Previously this stanza was duplicated in sum and prod.

Returning a copy looks reasonable, too.



def prod(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
dtype: Optional[DType] = None,
keepdims: bool = False,
**kwargs) -> Array:
ndim = x.ndim

# https://github.com/pytorch/pytorch/issues/29137. Separate from the logic
# below because it still needs to upcast.
if axis == ():
if dtype is None:
# We can't upcast uint8 according to the spec because there is no
# torch.uint64, so at least upcast to int64 which is what sum does
# when axis=None.
if x.dtype in [torch.int8, torch.int16, torch.int32, torch.uint8]:
return x.to(torch.int64)
return x.clone()
return x.to(dtype)

return _sum_prod_no_axis(x, dtype)
# torch.prod doesn't support multiple axes
# (https://github.com/pytorch/pytorch/issues/56586).
if isinstance(axis, tuple):
Expand All @@ -330,7 +322,7 @@ def prod(x: Array,
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.prod(x, dtype=dtype, **kwargs)
res = _axis_none_keepdims(res, ndim, keepdims)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res

return torch.prod(x, axis, dtype=dtype, keepdims=keepdims, **kwargs)
Expand All @@ -343,25 +335,14 @@ def sum(x: Array,
dtype: Optional[DType] = None,
keepdims: bool = False,
**kwargs) -> Array:
ndim = x.ndim

# https://github.com/pytorch/pytorch/issues/29137.
# Make sure it upcasts.
if axis == ():
if dtype is None:
# We can't upcast uint8 according to the spec because there is no
# torch.uint64, so at least upcast to int64 which is what sum does
# when axis=None.
if x.dtype in [torch.int8, torch.int16, torch.int32, torch.uint8]:
return x.to(torch.int64)
return x.clone()
return x.to(dtype)

return _sum_prod_no_axis(x, dtype)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.sum(x, dtype=dtype, **kwargs)
res = _axis_none_keepdims(res, ndim, keepdims)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res

return torch.sum(x, axis, dtype=dtype, keepdims=keepdims, **kwargs)
Expand All @@ -372,7 +353,7 @@ def any(x: Array,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
**kwargs) -> Array:
ndim = x.ndim

if axis == ():
return x.to(torch.bool)
# torch.any doesn't support multiple axes
Expand All @@ -384,7 +365,7 @@ def any(x: Array,
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.any(x, **kwargs)
res = _axis_none_keepdims(res, ndim, keepdims)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res.to(torch.bool)

# torch.any doesn't return bool for uint8
Expand All @@ -396,7 +377,7 @@ def all(x: Array,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
**kwargs) -> Array:
ndim = x.ndim

if axis == ():
return x.to(torch.bool)
# torch.all doesn't support multiple axes
Expand All @@ -408,7 +389,7 @@ def all(x: Array,
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.all(x, **kwargs)
res = _axis_none_keepdims(res, ndim, keepdims)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res.to(torch.bool)

# torch.all doesn't return bool for uint8
Expand Down
Loading