[Feature Request] key_equal, copy_key, zero_key in dict is slower than direct assignment if key type is primitive

[dlee992]

Hi, I noticed an unoptimized situation that key_equal, copy_key, zero_key in dict are slower than direct assignment if key type is primitive. The root cause is if key_type doesn't contain meminfo, then key_equal will rollback to using memcmp, which is pretty slow compared to directly this_key == an_integer. Other two functions will rollback to using memcpy, which is also slow.

numba/numba/cext/dictobject.c

Line 448 in a060559

return memcmp(lhs, rhs, dk->key_size) == 0;

numba/numba/cext/dictobject.c

Line 434 in a060559

memcpy(dst, src, dk->key_size);

I think we can do more things in this intrinsic function, which includes generation of corresponding functions (i.e., key_equal, copy_key, zero_key) for primitive types. I have already tested this in an internal use-case, this optimization can boost performance at least 5%~10% if using numba typed.dict heavily (i.e., lots of dict lookup, insert operations)

numba/numba/typed/dictobject.py

Line 264 in a060559

def _dict_set_method_table(typingctx, dp, keyty, valty):

[guilhermeleobas]

Hi @dlee992, do you have a reproducer for one of the functions you mentioned?

[dlee992]

Hi, @guilhermeleobas.

Sure, I should provide an example to show it, I will give one after I provide a corresponding PR, then we can compare the performance changes with and without this optimization.

But I can ensure this situation is real for heavy usage of numba typed.Dict.

[github-actions]

This issue is marked as stale as it has had no activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with any updates and confirm that this issue still needs to be addressed.