9prady9 commented on a change in pull request #11674:
URL: https://github.com/apache/arrow/pull/11674#discussion_r748163500
##########
File path: cpp/src/arrow/util/bitmap_ops.cc
##########
@@ -249,59 +249,92 @@ void AlignedBitmapOp(const uint8_t* left, int64_t
left_offset, const uint8_t* ri
left += left_offset / 8;
right += right_offset / 8;
out += out_offset / 8;
+ uint64_t outPopCount = 0;
for (int64_t i = 0; i < nbytes; ++i) {
out[i] = op(left[i], right[i]);
+ if (ComputeNewValidityCount) {
+ outPopCount += BitUtil::kBytePopcount[out[i]];
Review comment:
> The problem is that v is calculated but never used. So compiler will
simply delete all the code calculating it.
> Change the function to let it return v, the generated code will be totally
different.
Makes sense. After playing with it for few minutes, saving result to
register and then using that for other two calculations is what letting
compiler vectorize the loop on all three compilers consistently.
- clang - https://godbolt.org/z/sKvodGzbr
- gcc - https://godbolt.org/z/Kdc38P7fx
- msvc - https://godbolt.org/z/n1so9PbG3
In any case, eventually I dropped the above approach.
I had a version of CountSetBits earlier that was doing popcount on 64-bit
words. But I removed that in favour of lookup during the bitmap op assuming
lookup might be faster. It turned out just doing popcount finally on the whole
array is more efficient than clubbing the pop-count with bitmapOp .
Please check the updated description for new numbers
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
