Use SIMD instructions for bitwise operations (gcc specific). #20
Conversation
* Avoid 'case' overhead by generating BITWISE_FUNC vs. BITARRAY_FUNC.
* Change type of zero and one to (unsigned char) due to compiler complaining of
overflows on '-pedantic'.
This resulted in a ~10x speedup on large bitarrays for me using the following
test.
```python
import bitarray
import timeit
a = bitarray.bitarray(50000)
b = bitarray.bitarray(50000)
def test_and():
global a
a &= b
def test_or():
global a
a |= b
def test_xor():
global a
a ^= b
print timeit.timeit("test_and()", "from __main__ import test_and")
print timeit.timeit("test_or()", "from __main__ import test_or")
print timeit.timeit("test_xor()", "from __main__ import test_xor")
```
```
upstream master:
20.3912520409
20.6214001179
20.5252711773
with this patch:
2.11912703514
2.14890694618
2.1437420845
```
About memcpy usage in simd_v16uc_op:
I found that memcpy did the most clever thing in most cases when inspecting
compiler output.
On my system it uses movdqa to copy memory to and from xmm registers.
|
This looks really promising. Several operations need to be faster with bitarray, and as long as everything passes tests, I support this PR. |
|
Hey @udoprog, this repo has been silent for quite a while, and we thus created a fork at https://github.com/diamondman/bitarray/. If you are interested, please feel free to transplant your pull request to the forked repo, we would be very happy to begin the code review and merge into bitarray before pushing out a new release. If you do not have the time to do so, we would kindly ask your permission to do the PR transfer our-self. If you could ping back in the next couple of days, one way or the other, that would be great. Many thanks! |
|
Thank you for your PR, and sorry for the long response time. I recently used uint64 integers to optimize bitwise operations, see #133, in a non-gcc specific way. |
4 participants
overflows on '-pedantic'. Also add 'static' to help the compiler out more.
This resulted in a ~10x speedup on large bitarrays for me using the following
test.
About memcpy usage in simd_v16uc_op:
I found that memcpy did the most clever thing in most cases when inspecting
compiler output.
On my system it uses movdqa to copy memory to and from xmm registers.