Commit message (Collapse) | Author | Age | Files | Lines | |
---|---|---|---|---|---|
* | Move ExponentialBackoff to vespajlib | Martin Polden | 2024-03-05 | 1 | -0/+47 |
| | |||||
* | Add fluent builder | Bjørn Christian Seime | 2024-02-15 | 1 | -1/+1 |
| | |||||
* | Proxy endpoint certificate secrets through EndpointCertificateSecretStore | Morten Tokle | 2024-02-13 | 1 | -0/+17 |
| | |||||
* | - Drop the optimization from long to int and just use long throughout. | Henning Baldersheim | 2024-02-10 | 11 | -120/+101 |
| | | | | - That enables that numeric hashes are also handled efficiently, without resorting to strings. | ||||
* | Support embedding into rank 3 tensors | Jon Bratseth | 2024-02-02 | 1 | -0/+1 |
| | |||||
* | Reverse the problem to match the interface of TensorAddress.partialCopy | Henning Baldersheim | 2024-02-01 | 1 | -6/+2 |
| | |||||
* | Update abi. | Henning Baldersheim | 2024-02-01 | 1 | -1/+1 |
| | |||||
* | - Use numericLabel over label for address manipulation. | Henning Baldersheim | 2024-02-01 | 6 | -51/+29 |
| | | | | - Only use label when actual string representation is needed. | ||||
* | Limit optimization to bound types, which is the prevalent kind. | Henning Baldersheim | 2024-01-31 | 1 | -23/+27 |
| | |||||
* | - Make separate path for reduction of indexed tensors. | Henning Baldersheim | 2024-01-31 | 2 | -6/+74 |
| | | | | | | | - This brough down reduce time from 60s to 3.7s or a factor of 16x for the splade embedder if using reduce. This is ontop of the 1.5x improvement with https://github.com/vespa-engine/vespa/pull/30112 - A reduce based splade embedder now uses 8.0s vs 7.0s for a custom rolled out version, translated to 14% overhead. Please enter the commit message for your changes. Lines starting | ||||
* | Use already optimize TensorAddress.partialCopy to reduce dimensions. | Henning Baldersheim | 2024-01-30 | 1 | -9/+2 |
| | |||||
* | Add copyright and minor cleanup. | Henning Baldersheim | 2024-01-30 | 1 | -2/+2 |
| | |||||
* | - 0 prefixed numbers should not be considered numeric." | Henning Baldersheim | 2024-01-30 | 1 | -0/+1 |
| | |||||
* | Minor cleanup | Jon Bratseth | 2024-01-28 | 18 | -84/+159 |
| | |||||
* | Make tensor addresses integer based instead of as strings. | Henning Baldersheim | 2024-01-25 | 22 | -298/+630 |
| | | | | | Positive numbers are direct indexes, while strings that does not represent numbers are enumerated and represented with negative integers. | ||||
* | Add TensorAdressAny.java | Henning Baldersheim | 2024-01-25 | 2 | -0/+21 |
| | |||||
* | Support longhasher | Henning Baldersheim | 2024-01-25 | 1 | -0/+17 |
| | |||||
* | - Replace usage of putIfAbsent and manual optimizes get-if-not-put with ↵ | Henning Baldersheim | 2024-01-22 | 4 | -16/+7 |
| | | | | computeIfAbsent. | ||||
* | Make address generation more efficient, also prepare for presizing during build. | Henning Baldersheim | 2024-01-22 | 1 | -23/+37 |
| | |||||
* | Extract label once. | Henning Baldersheim | 2024-01-22 | 2 | -3/+3 |
| | |||||
* | - Using optional in the inner loop is often very inefficient. It bloats the code | Henning Baldersheim | 2024-01-22 | 2 | -3/+10 |
| | | | | and most of the time hampers the compilers abilitity to inline and optimize. | ||||
* | Make tests independent of hash function | Henning Baldersheim | 2024-01-21 | 1 | -1/+2 |
| | |||||
* | Use lz4-java and xxhash32 | Henning Baldersheim | 2024-01-21 | 1 | -4/+16 |
| | |||||
* | - Add and use getAsDouble method returning a Double object. It behaves ↵ | Henning Baldersheim | 2024-01-21 | 6 | -8/+37 |
| | | | | | | | similar to Map.get(key). null indicates no value present. Then you avoid 2 lookups to first check if a value is present, and then to fetch that value. This does wonders when it is backed by a map and hashCode/equals are relatively expensive. | ||||
* | Merge pull request #29996 from vespa-engine/balder/do-sanity-checking-in-builder | Henning Baldersheim | 2024-01-21 | 3 | -79/+71 |
|\ | | | | | - Avoid expensive iteration with hash lookups just for sanity checkin… | ||||
| * | - Avoid expensive iteration with hash lookups just for sanity checking in ↵ | Henning Baldersheim | 2024-01-21 | 3 | -79/+71 |
| | | | | | | | | | | | | private MixedTensor constructor. - Refactor so there is less need for sanity checks, and do them in the buiulder instead. | ||||
* | | Revert "Balder/cached xxh3 hash" | Henning Baldersheim | 2024-01-21 | 3 | -37/+7 |
|/ | |||||
* | - Use get(key) instead of containsKey to avoid extra lookup. | Henning Baldersheim | 2024-01-21 | 1 | -25/+15 |
| | | | | - Enough to use only array[long] and convert directly. No need to create temporary TensorAddress. | ||||
* | Cache size of intersected sets, as they are recomputed every time. | Henning Baldersheim | 2024-01-21 | 1 | -5/+7 |
| | |||||
* | Merge pull request #29993 from ↵ | Jon Bratseth | 2024-01-21 | 9 | -33/+54 |
|\ | | | | | | | | | vespa-engine/balder/precompute-type-related-information-once - Extract dimension names in a set to avoid recomputing it in dimensi… | ||||
| * | Make the TensorType.hasXX public and use them other places too. | Henning Baldersheim | 2024-01-21 | 6 | -20/+17 |
| | | |||||
| * | - Extract dimension names in a set to avoid recomputing it in ↵ | Henning Baldersheim | 2024-01-21 | 4 | -16/+40 |
| | | | | | | | | | | | | | | dimensionNames() mathod. - Also extract some information about what kind of dimensions a tensor has construcion time. This avoids streamin through all diemsion later on, and keeps this local in the TensorType. | ||||
* | | of => withSeed | Henning Baldersheim | 2024-01-21 | 2 | -2/+2 |
| | | |||||
* | | Must handle null dimensions. | Henning Baldersheim | 2024-01-21 | 1 | -1/+4 |
| | | |||||
* | | Use a LinkedHashMap to make json tests independant of hash function. | Henning Baldersheim | 2024-01-21 | 1 | -1/+2 |
| | | |||||
* | | - Use xxh3 hash for better hashing, and cache the hash value. | Henning Baldersheim | 2024-01-21 | 2 | -6/+32 |
|/ | |||||
* | Rename getIndex => getDirectIndex | Henning Baldersheim | 2024-01-20 | 2 | -3/+3 |
| | |||||
* | Add a class for assist efficient traversal of dimensions in an IndexedTensor. | Henning Baldersheim | 2024-01-19 | 2 | -0/+42 |
| | |||||
* | - Move computation of productOfDimensionsAfter to DimensionSizes. | Henning Baldersheim | 2024-01-18 | 2 | -47/+45 |
| | | | | - And then precompute them construction time. | ||||
* | Construct array right away instead of going via a single element list and ↵ | Henning Baldersheim | 2024-01-18 | 3 | -4/+4 |
| | | | | the java stream api. | ||||
* | Skip copying of array when there is no reuse possible. | Henning Baldersheim | 2024-01-18 | 5 | -10/+16 |
| | |||||
* | - Make an tensor.impl package that can be used from other tensor packages ↵ | Henning Baldersheim | 2024-01-18 | 5 | -89/+125 |
| | | | | | | | | | allowing bypass of defensive strategies in the public interfaces. - Move private static inner classes TensorAddress.NumericTensorAddress and TensorAddress.StringTensorAddress to tensor.impl package. - Use the StringTensorAddress.of from Reduce to avoid defensive array copy. | ||||
* | - Add sizeAsInt to allow for safe cast from long to int of the size of a tensor. | Henning Baldersheim | 2024-01-18 | 7 | -12/+37 |
| | |||||
* | Update vespajlib/src/main/java/com/yahoo/tensor/functions/Reduce.java | Henning Baldersheim | 2024-01-18 | 1 | -1/+1 |
| | | | | | Unify whitespace Co-authored-by: Jon Bratseth <bratseth@gmail.com> | ||||
* | Update vespajlib/src/main/java/com/yahoo/tensor/functions/Reduce.java | Henning Baldersheim | 2024-01-18 | 1 | -1/+1 |
| | | | | | Unify whitespace Co-authored-by: Jon Bratseth <bratseth@gmail.com> | ||||
* | Update vespajlib/src/main/java/com/yahoo/tensor/functions/Reduce.java | Henning Baldersheim | 2024-01-18 | 1 | -1/+1 |
| | | | | | Unify whitespace Co-authored-by: Jon Bratseth <bratseth@gmail.com> | ||||
* | Update vespajlib/src/main/java/com/yahoo/tensor/functions/Reduce.java | Henning Baldersheim | 2024-01-18 | 1 | -1/+1 |
| | | | | | Unify whitespace Co-authored-by: Jon Bratseth <bratseth@gmail.com> | ||||
* | Update vespajlib/src/main/java/com/yahoo/tensor/functions/Reduce.java | Henning Baldersheim | 2024-01-18 | 1 | -1/+1 |
| | | | | | Unify whitespace Co-authored-by: Jon Bratseth <bratseth@gmail.com> | ||||
* | Reverse the problem from indexes-to-remove to indexes-to-keep. Then you ↵ | Henning Baldersheim | 2024-01-17 | 1 | -13/+19 |
| | | | | avoid hash lookup in inner loop. | ||||
* | Create set of indexes to remove once. | Henning Baldersheim | 2024-01-17 | 2 | -5/+14 |
| |