Skip to content

PERF: BlockManager.equals blockwise #35357

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.

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 13 commits into from
Aug 7, 2020
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
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
33 changes: 13 additions & 20 deletions pandas/core/internals/managers.py
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@
get_block_type,
make_block,
)
from pandas.core.internals.ops import operate_blockwise
from pandas.core.internals.ops import blockwise_all, operate_blockwise

# TODO: flexible with index=None and/or items=None

Expand Down Expand Up @@ -1422,32 +1422,25 @@ def equals(self, other: "BlockManager") -> bool:
if not all(ax1.equals(ax2) for ax1, ax2 in zip(self_axes, other_axes)):
return False

if self.ndim == 1:
# For SingleBlockManager (i.e.Series)
if other.ndim != 1:
return False
left = self.blocks[0].values
right = other.blocks[0].values
def blk_func(left, right):
Copy link
Contributor

Choose a reason for hiding this comment

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

can you make this module left

Copy link
Contributor

Choose a reason for hiding this comment

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

can you do this; make this a module level function

if not is_dtype_equal(left.dtype, right.dtype):
return False
elif isinstance(left, ExtensionArray):
return left.equals(right)
elif isinstance(right, ExtensionArray):
return False
else:
return array_equivalent(left, right)
return array_equivalent(left, right, dtype_equal=True)

for i in range(len(self.items)):
# Check column-wise, return False if any column doesn't match
left = self.iget_values(i)
right = other.iget_values(i)
if not is_dtype_equal(left.dtype, right.dtype):
if self.ndim == 1:
# For SingleBlockManager (i.e.Series)
if other.ndim != 1:
return False
elif isinstance(left, ExtensionArray):
if not left.equals(right):
return False
else:
if not array_equivalent(left, right, dtype_equal=True):
return False
return True
left = self.blocks[0].values
right = other.blocks[0].values
return blk_func(left, right)

return blockwise_all(self, other, blk_func)

def unstack(self, unstacker, fill_value) -> "BlockManager":
"""
Expand Down
47 changes: 31 additions & 16 deletions pandas/core/internals/ops.py
Original file line number Diff line number Diff line change
Expand Up @@ -9,13 +9,10 @@
from pandas.core.internals.blocks import Block # noqa:F401


def operate_blockwise(
left: "BlockManager", right: "BlockManager", array_op
) -> "BlockManager":
def _iter_block_pairs(left: "BlockManager", right: "BlockManager"):
Copy link
Contributor

Choose a reason for hiding this comment

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

can you type the output

# At this point we have already checked the parent DataFrames for
# assert rframe._indexed_same(lframe)

res_blks: List["Block"] = []
for n, blk in enumerate(left.blocks):
locs = blk.mgr_locs
blk_vals = blk.values
Expand All @@ -34,21 +31,31 @@ def operate_blockwise(
right_ea = not isinstance(rblk.values, np.ndarray)

lvals, rvals = _get_same_shape_values(blk, rblk, left_ea, right_ea)
yield lvals, rvals, locs, left_ea, right_ea, rblk
Copy link
Contributor

Choose a reason for hiding this comment

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

maybe yield a NamedTuple?

Copy link
Member Author

Choose a reason for hiding this comment

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

IIRC there are issues with NamedTuple construction performance

Copy link
Member

Choose a reason for hiding this comment

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

maybe wait a couple of days and use a dataclass.

Copy link
Contributor

Choose a reason for hiding this comment

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

IIRC there are issues with NamedTuple construction performance

really? can you point to something that shows this

Copy link
Member Author

Choose a reason for hiding this comment

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

May be out of date but I think this is where i got the impression that NamedTuple has perf issues: https://lwn.net/Articles/731423/

Copy link
Contributor

Choose a reason for hiding this comment

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

i prefer readability here
unless it's way slower


res_values = array_op(lvals, rvals)
if left_ea and not right_ea and hasattr(res_values, "reshape"):
res_values = res_values.reshape(1, -1)
nbs = rblk._split_op_result(res_values)

# Assertions are disabled for performance, but should hold:
# if right_ea or left_ea:
# assert len(nbs) == 1
# else:
# assert res_values.shape == lvals.shape, (res_values.shape, lvals.shape)
def operate_blockwise(
left: "BlockManager", right: "BlockManager", array_op
) -> "BlockManager":
# At this point we have already checked the parent DataFrames for
# assert rframe._indexed_same(lframe)

res_blks: List["Block"] = []
for lvals, rvals, locs, left_ea, right_ea, rblk in _iter_block_pairs(left, right):
res_values = array_op(lvals, rvals)
if left_ea and not right_ea and hasattr(res_values, "reshape"):
res_values = res_values.reshape(1, -1)
nbs = rblk._split_op_result(res_values)

_reset_block_mgr_locs(nbs, locs)
# Assertions are disabled for performance, but should hold:
# if right_ea or left_ea:
# assert len(nbs) == 1
# else:
# assert res_values.shape == lvals.shape, (res_values.shape, lvals.shape)

res_blks.extend(nbs)
_reset_block_mgr_locs(nbs, locs)

res_blks.extend(nbs)

# Assertions are disabled for performance, but should hold:
# slocs = {y for nb in res_blks for y in nb.mgr_locs.as_array}
Expand Down Expand Up @@ -85,7 +92,7 @@ def _get_same_shape_values(
# Require that the indexing into lvals be slice-like
assert rblk.mgr_locs.is_slice_like, rblk.mgr_locs

# TODO(EA2D): with 2D EAs pnly this first clause would be needed
# TODO(EA2D): with 2D EAs only this first clause would be needed
if not (left_ea or right_ea):
lvals = lvals[rblk.mgr_locs.indexer, :]
assert lvals.shape == rvals.shape, (lvals.shape, rvals.shape)
Expand All @@ -102,3 +109,11 @@ def _get_same_shape_values(
rvals = rvals[0, :]

return lvals, rvals


def blockwise_all(left: "BlockManager", right: "BlockManager", op) -> bool:
for lvals, rvals, _, _, _, _ in _iter_block_pairs(left, right):
Copy link
Contributor

Choose a reason for hiding this comment

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

can you add a doc-string (also using a NamedTuple makes this easier to read here)

res = op(lvals, rvals)
if not res:
return False
return True