Gaussian Elimination #4
Conversation
|
LGTM. It might be nice to add a test for an unsolvable system. |
| # Using gmpy to find the index of the least significant bit. Sources: | ||
| # https://stackoverflow.com/questions/5520655/return-index-of-least-significant-bit-in-python | ||
| # https://stackoverflow.com/questions/12078277/is-this-a-bug-in-gmpy2-or-am-i-mad | ||
| index_of_first_set_bit_in_chunk = gmpy2.bit_scan1(gmpy2.mpz(chunk)) |
There was a problem hiding this comment.
This is do-able via a few bit shift ops - is gmpy2 faster? If not, then maybe we can remove the gmpy dependency.
There was a problem hiding this comment.
I haven't benchmarked it but the stack overflow post I linked claims gmpy is 4x faster than using ctypes. not sure if that uses bit ops but we can revisit this later if needed
| # TODO: np.where searches the whole array, and doesnt stop at the first | ||
| # sight of nonzero, can we optimize this? | ||
| equation = self._equations[equation_id] | ||
| first_nonzero_chunk_id = np.where(equation != 0)[0][0] |
There was a problem hiding this comment.
In LazyGaussianElimination.py, we use:
for chunk_id, chunk in enumerate(equation):
if flag > 0:
break
lsb_index = chunk_id * self._num_variables_per_chunk + index_of_first_set_bit_in_chunk
Let's check for performance on equations with ~1 million entries - it's a question of whether Python loops with early termination win against C loops without early termination.
There was a problem hiding this comment.
sounds good. ill keep the TODO so we can profile this if needed
No description provided.