PERF/REF: Check use of numexpr earlier in the DataFrame operation #41122
Conversation
jorisvandenbossche
added
Refactor
| ---------- | ||
| op : operator | ||
| size : int | ||
| dtypes : list |
| else: | ||
| use_numexpr = expressions.can_use_numexpr(func, None, None, right) | ||
|
|
||
| array_op = ops.get_array_op(func, use_numexpr=use_numexpr) |
There was a problem hiding this comment.
is the get_array_op on L6832 still needed?
|
|
||
| if isinstance(self._mgr, ArrayManager): | ||
| use_numexpr = expressions.can_use_numexpr( | ||
| func, self.shape[1], (right.dtype,), None |
There was a problem hiding this comment.
does self.shape[1] correspond to the size of some array?
| else: | ||
| use_numexpr = expressions.can_use_numexpr( | ||
| func, None, (right.dtype,), None | ||
| ) |
There was a problem hiding this comment.
these ~8 lines (x4) seem like a reasonable cope for an AM/BM method
|
|
||
|
|
||
| def _na_arithmetic_op(left, right, op, is_cmp: bool = False): | ||
| def _na_arithmetic_op(left, right, op, is_cmp: bool = False, use_numexpr=True): |
There was a problem hiding this comment.
can use_numexpr be keyword-only
|
any estimates for perf impact? |
|
This pull request is stale because it has been open for thirty days with no activity. Please update or respond to this comment if you're still interested in working on this. |
#41161 and #40463 have been merged. can you address @jbrockmendel comments. |
| Parameters | ||
| ---------- | ||
| op : operator | ||
| size : int |
|
|
||
| # required min elements (otherwise we are adding overhead) | ||
| if a.size > _MIN_ELEMENTS: | ||
| if isinstance(b, str): |
There was a problem hiding this comment.
the place where this is moved from has a comment # can never use numexpr. not a huge loss since the comment isn't that informative, but at some point we should get a helpful comment about why it cant be used
|
Thanks for the PR but appears to have gone stale. Closing to clear the queue, but feel free to reopen if you find the time to revisit this PR and the comments. |
4 participants
This PR tries to gather some checks about whether numexpr can be used or not which can be done upfront on the DataFrame level, before dispatching to the array_ops.
For example, for a scalar right operand, we can check it only once if it's compatible with numexpr usage. Or for ArrayManager, we can check the size constraint once upfront (since each array has the same size). For those cases, this avoids overhead in
expressions.evaluate.To this end, I created a new
can_use_numexprthat starts to gather some checks (there might be more that can be added), call this up front inDataFrame._dispatch_frame_op(this could IMO be bit cleaner with #39772, but the if/else checks inside_dispatch_frame_opwork as well). And then the result ofcan_use_numexpris passed through to the array_ops.