summaryrefslogtreecommitdiffstats
path: root/eval
Commit message (Collapse)AuthorAgeFilesLines
* Add dimension memory cost too.Henning Baldersheim2023-10-242-7/+8
|
* Document memory footprint of vespalib:.eval::FastValueHenning Baldersheim2023-10-241-3/+27
|
* ValueType copy constructor/operatorHenning Baldersheim2023-10-243-8/+11
|
* mallinfo is linux specific.Tor Egge2023-10-191-0/+8
|
* use same variable nameArne Juul2023-10-191-1/+1
|
* Provide both number from malloc and from vm to compare. Should most likely ↵Henning Baldersheim2023-10-191-8/+38
| | | | use max of these go best estimate model cost.
* Report peak memory usage.Henning Baldersheim2023-10-182-10/+9
|
* Simplify include of onnxruntime header file.Tor Egge2023-10-151-4/+0
|
* Update copyrightJon Bratseth2023-10-09467-467/+467
|
* Add missing includes, avoid shadow warning and skip including file notTor Egge2023-09-292-3/+5
| | | | | | present in llvm 17. Issues detected when compiling with clang++ 17 / libc++ 17 / llvm 17.
* Don't read from invalid file pointer (e.g. whenTor Egge2023-09-221-4/+6
| | | | /proc/self/status could not be opened)
* allow universal dot product optimizationHåvard Pettersen2023-09-141-1/+1
| | | | only used for cases where both inputs have mapped dimensions
* improve testing by verifying corner casesHåvard Pettersen2023-09-125-36/+78
|
* handle 'distinct' and 'single' flags using templatesHåvard Pettersen2023-09-086-153/+205
|
* also test with trivial dimensionsHåvard Pettersen2023-09-072-6/+4
|
* handle expanding reduceHåvard Pettersen2023-09-072-11/+97
| | | | more testing of corner cases
* Merge pull request #28413 from ↵Henning Baldersheim2023-09-064-162/+191
|\ | | | | | | | | vespa-engine/havardpe/avoid-making-new-value-index detect not having to make a new value index
| * detect not having to make a new value indexHåvard Pettersen2023-09-054-162/+191
| | | | | | | | | | | | - more benchmarking - drop expression descriptions - inline sparse plan execution
* | Merge pull request #28376 from vespa-engine/balder/minor-cleanupHenning Baldersheim2023-09-042-28/+57
|\ \ | | | | | | Unify and modernize code and layout
| * | Better namingHenning Baldersheim2023-09-041-9/+9
| | |
| * | Probe with optimize as in production.Henning Baldersheim2023-09-041-5/+3
| | |
| * | Keep oneliners for free functions.Henning Baldersheim2023-09-041-8/+4
| | |
| * | Report memory usage back when probing types.Henning Baldersheim2023-09-041-13/+48
| | |
| * | Unify and modernize code and layoutHenning Baldersheim2023-09-041-16/+16
| |/
* / Avoid operator<=> (libc++ std::vector limitation).Tor Egge2023-09-042-3/+7
|/
* benchmark universal dot product vs other optionsHåvard Pettersen2023-09-018-25/+336
|
* added universal dot productHåvard Pettersen2023-08-299-1/+250
| | | | note that optimization is not yet active in production
* Merge pull request #28156 from vespa-engine/havardpe/dense-join-reduce-planArne H Juul2023-08-2811-25/+372
|\ | | | | dense join reduce plan
| * dense join reduce planHåvard Pettersen2023-08-2511-25/+372
| | | | | | | | | | added ValueType::stride_of added nested_loop support for three parallel loops
* | Merge pull request #28122 from vespa-engine/arnej/onnx-1-15-1Arne H Juul2023-08-281-1/+1
|\ \ | |/ |/| use onnxruntime version 1.15.1
| * use onnxruntime version 1.15.1Arne Juul2023-08-231-1/+1
| |
* | added insert_subspace to fast valueHåvard Pettersen2023-08-242-0/+82
| |
* | sparse join reduce planHåvard Pettersen2023-08-237-0/+437
|/
* use common inlined code for low-level dot productsHåvard Pettersen2023-08-146-81/+44
|
* Add information about the incorrect spec causing the error.Henning Baldersheim2023-08-031-1/+3
|
* - Pack data closer to let config fit in 2 cache lines instead of 4.Henning Baldersheim2023-07-271-10/+10
| | | | | - Avoid plt indirection and allow more inlining of frequently called code. - Reapplication of #27646
* Revert "- Pack data closer to let config fit in 2 cache lines instead of 4."Tor Egge2023-07-141-10/+10
|
* Merge pull request #27655 from ↵Tor Egge2023-07-107-35/+101
|\ | | | | | | | | vespa-engine/arnej/defensive-constant-tensor-loader Arnej/defensive constant tensor loader
| * allow integer labels for literal format, but add strict checking ofArne Juul2023-07-067-35/+101
| | | | | | | | addresses to avoid crashing into asserts in CreateValueFromTensorSpec
* | Revert "Add DenseValuesDecoder constructor (needed by AppleClang)."Tor Egge2023-07-101-5/+0
| |
* | Add DenseValuesDecoder constructor (needed by AppleClang).Tor Egge2023-07-061-0/+5
| |
* | - Pack data closer to let config fit in 2 cache lines instead of 4.Henning Baldersheim2023-07-061-10/+10
|/ | | | - Avoid plt indirection and allow more inlining of frequently called code.
* handle even more JSON formats for tensorsArne Juul2023-07-036-22/+145
|
* style fixesArne Juul2023-06-291-12/+12
|
* add MixedL2Distance optimizerArne Juul2023-06-287-1/+245
|
* Eliminate inline warning in dense multi matmul function unit test.Tor Egge2023-06-062-2/+4
|
* allow array of numbers at top level; unit test short-short formsArne Juul2023-05-196-14/+18
|
* allow short-form JSON for 1-d constantsArne Juul2023-05-164-12/+101
|
* Move FastValueView to separate files.Tor Egge2023-02-232-11/+25
|
* Add noexcept and some constexprHenning Baldersheim2023-02-111-23/+23
|