aboutsummaryrefslogtreecommitdiffstats
path: root/searchlib
Commit message (Collapse)AuthorAgeFilesLines
* Merge pull request #31336 from ↵Tor Egge2024-05-295-66/+114
|\ | | | | | | | | vespa-engine/toregge/factor-out-first-phase-rescorer-from-hit-collector Factor out FirstPhaseRescorer from HitCollector.
| * Pass docid to rescorers.Tor Egge2024-05-292-4/+5
| |
| * Factor out FirstPhaseRescorer from HitCollector.Tor Egge2024-05-295-66/+113
| |
* | - Add is_search_multi_threaded to proton::matching::BlueprintBuilder for ↵Henning Baldersheim2024-05-291-1/+1
| | | | | | | | | | | | consistency and readability. - Fix CreateBlueprintVisitorHelper to actually detect singlethreaded usage.
* | - Separate priority Q functionality into a separate base class ↵Henning Baldersheim2024-05-2810-49/+83
|/ | | | | | | WeakAndPriorityQueue. - Keep thread safety in the SharedWeakAndpriorityQueue. - Only use thread safe variant if more than 1 thread in the search bundle.
* Use shared heapHenning Baldersheim2024-05-282-9/+16
|
* Merge pull request #31282 from ↵Henning Baldersheim2024-05-2820-192/+233
|\ | | | | | | | | vespa-engine/balder/wire-parallel-heap-to-simple-weakand - Wire in a common shared heap for WeakAnd.
| * Prepare heap correctlyHenning Baldersheim2024-05-244-10/+10
| |
| * - Wire in a common shared heap for WeakAnd.Henning Baldersheim2024-05-2318-185/+226
| |
* | Add firstPhaseRank feature (step 1, not yet integrated with match loop).Tor Egge2024-05-2411-0/+379
|/
* Pass all params in constructor and make members const.Henning Baldersheim2024-05-223-20/+17
|
* Split ParallellWandMatchParams in 2.Henning Baldersheim2024-05-224-26/+36
|
* It is already a unique_ptrHenning Baldersheim2024-05-211-10/+7
|
* Use std::forward recommended by clion.Henning Baldersheim2024-05-212-14/+16
|
* libc++ 18 still needs workarounds.Tor Egge2024-05-211-1/+1
|
* Use __attribute__((noinline)) instead of glibc macro __attribute_noinline__Tor Egge2024-05-202-4/+4
|
* Remove empty comment.Tor Egge2024-05-161-1/+0
|
* Add second phase feature.Tor Egge2024-05-166-0/+138
|
* Avoid unresolved symbols when compiling without optimization.Tor Egge2024-05-161-3/+3
|
* Merge pull request #31210 from vespa-engine/balder/faster-int8floatHenning Baldersheim2024-05-167-23/+29
|\ | | | | Speed up dotproduct for int8.
| * Now there is preliminary native support for Int8Float too.Henning Baldersheim2024-05-151-5/+8
| |
| * Use a valuespace representable by both 8,16,23 and 64 bit floats to easier ↵Henning Baldersheim2024-05-151-3/+3
| | | | | | | | verify correctness.
| * Speed up dotproduct for int8.Henning Baldersheim2024-05-156-15/+18
| |
* | Merge pull request #31209 from ↵Håvard Pettersen2024-05-151-1/+1
|\ \ | |/ |/| | | | | vespa-engine/geirst/attribute-weighted-set-blueprint-flow-stats Update flow stats to reflect that reverse hash lookup is always used …
| * Update flow stats to reflect that reverse hash lookup is always used when ↵Geir Storli2024-05-151-1/+1
| | | | | | | | non-strict.
* | Use explicit static_castHenning Baldersheim2024-05-151-3/+3
| |
* | Speed up bfloat16 to float conversionHenning Baldersheim2024-05-141-4/+17
|/
* Merge pull request #31200 from vespa-engine/balder/add-distance-benchmarkHenning Baldersheim2024-05-1416-38/+167
|\ | | | | Add benchmark for distance functions
| * Avoid possible compiler optimisation.Henning Baldersheim2024-05-141-18/+32
| |
| * Use use method without limit which already use hw optimized instructions.Henning Baldersheim2024-05-141-8/+1
| |
| * Add benchmark for distance functionsHenning Baldersheim2024-05-1416-30/+152
| |
* | Merge pull request #31201 from ↵Geir Storli2024-05-146-107/+14
|\ \ | | | | | | | | | | | | vespa-engine/havardpe/remove-actual-cost-estimates remove code to estimate actual cost
| * | remove code to estimate actual costHåvard Pettersen2024-05-146-107/+14
| | |
* | | Merge pull request #31199 from ↵Geir Storli2024-05-141-2/+2
|\ \ \ | |_|/ |/| | | | | | | | vespa-engine/toregge/adjust-invalid-feature-name-messages-when-setting-up-first-phase-and-second-phrase-rank-feature Adjust invalid feature name message when setting up first phase and
| * | Adjust invalid feature name message when setting up first phase andTor Egge2024-05-141-2/+2
| | | | | | | | | | | | second phase rank feature.
* | | Drop short circuit optimisation as it prevents vectorisation.Henning Baldersheim2024-05-132-3/+2
|/ /
* | Merge pull request #31184 from vespa-engine/balder/unify-codeHenning Baldersheim2024-05-132-29/+19
|\ \ | | | | | | Unify code
| * | Unify codeHenning Baldersheim2024-05-132-29/+19
| | |
* | | Use an assertless typify in innerloopHenning Baldersheim2024-05-131-1/+1
| |/ |/|
* | Avoid assert in inner loop.Henning Baldersheim2024-05-131-1/+0
|/
* Merge pull request #31169 from ↵Geir Storli2024-05-102-9/+14
|\ | | | | | | | | vespa-engine/toregge/use-btree-iterator-move-constructor Use btree iterator move constructor.
| * Use btree iterator move constructor.Tor Egge2024-05-102-9/+14
| |
* | use least common term to drive strict phrase searchHåvard Pettersen2024-05-101-6/+4
|/ | | | | | | | | | stop using beginId since it only makes sense for the first (or full) docid space remove misleading assert. We could still end up with an infinite loop without triggering the assert if the driving term matched the next document. Also, this assert is not used with other iterators that require strict children.
* run non-strict source blenderHåvard Pettersen2024-05-071-30/+44
|
* Remove obsolete workaround for indexing exact-match fields.Tor Egge2024-05-062-3/+1
|
* Merge pull request #31128 from ↵v8.340.17Geir Storli2024-05-063-24/+77
|\ | | | | | | | | vespa-engine/havardpe/source-selector-iterator-benchmark baseline source blender integration in benchmark
| * baseline source blender integration in benchmarkHåvard Pettersen2024-05-063-24/+77
| |
* | Adjust down the estimate when its unknown and avoid WhiteListBlueprint being ↵Geir Storli2024-05-063-8/+15
|/ | | | | | | strict. In cases where only attributes without fast-search are searched with e.g. an OR, this change avoids sorting the WhiteListBlueprint first under the root AND, which is very costly to evaluate.
* test and adjust some stuffHåvard Pettersen2024-05-036-7/+52
|
* when is actual non-strict array lookup fasterHåvard Pettersen2024-05-021-11/+27
| | | | ... than forced strict btree posting list