Fix IVFPQFastScan decode function #3312
Conversation
|
This pull request was exported from Phabricator. Differential Revision: D55264781 |
Summary: as the [#issue3258](#3258) mentioned, the IVFPQFastScan should have same decoding result as IVFPQ. However, current result is not as expected. In this PR/Diff, we are going to fix the decoding function Differential Revision: D55264781
junjieqi
force-pushed
the
export-D55264781
branch
from
893e9cf
to
3abae7b
Compare
|
This pull request was exported from Phabricator. Differential Revision: D55264781 |
| pq.decode(bytes, x, n); | ||
| size_t coarse_size = coarse_code_size(); | ||
|
|
||
| #pragma omp parallel |
There was a problem hiding this comment.
parallel if (n > 1)
or, even better, if n is greater than a certain threshold
please check other sa_decode implementations for the appropriate threshold
There was a problem hiding this comment.
Hi @alexanderguzhva, thanks for the suggestion. I checked other sa_decode. They also don't have the condition. Try to understand it more. #pragma omp parallel if (n > 1) means we only parallel execution when n is greater than 1. I'm wondering how we should choose the threshold in general. Do we have any benchmark to run or other methods?
There was a problem hiding this comment.
no benchmarks, this is a magic constant. Typically, things like 100 or 1000 are used.
faiss/faiss/IndexScalarQuantizer.cpp
Line 186 in 0c96b0d
|
This pull request has been merged in 14b8af6. |
Summary: Pull Request resolved: facebookresearch#3312 as the [#issue3258](facebookresearch#3258) mentioned, the IVFPQFastScan should have same decoding result as IVFPQ. However, current result is not as expected. In this PR/Diff, we are going to fix the decoding function Reviewed By: mdouze Differential Revision: D55264781 fbshipit-source-id: dfdae9eabceadfc5a3ebb851930d71ce3c1c654d
Summary:
as the #issue3258 mentioned, the IVFPQFastScan should have same decoding result as IVFPQ. However, current result is not as expected.
In this PR/Diff, we are going to fix the decoding function
Differential Revision: D55264781