aboutsummaryrefslogtreecommitdiffstats
path: root/eval/src/tests/tensor/instruction_benchmark
Commit message (Collapse)AuthorAgeFilesLines
* use size literals in evalArne Juul2021-02-151-17/+18
|
* adjust param repo add options and remove seq_biasHåvard Pettersen2021-02-051-95/+95
| | | | GenSpec is now implicitly convertible to TensorSpec
* use GenSpec to generate test valuesHåvard Pettersen2021-02-011-136/+93
|
* forward or ignore index in relevant mixed tensor reduce casesHåvard Pettersen2021-01-191-0/+8
|
* forward index for concat of mixed tensor with dense tensorArne Juul2021-01-151-0/+6
|
* Merge pull request #15764 from vespa-engine/arnej/move-dense-optimizersArne H Juul2020-12-091-1/+0
|\ | | | | Arnej/move dense optimizers
| * move to vespalib::eval namespaceArne Juul2020-12-091-1/+0
| |
* | ghost support in instruction benchmarkHåvard Pettersen2020-12-092-8/+57
|/
* only factory in interpreted functionHåvard Pettersen2020-12-031-37/+36
|
* stop benchmarking old engineArne Juul2020-12-031-10/+6
|
* remove simple tensorHåvard Pettersen2020-12-021-1/+0
|
* GC unused codeArne Juul2020-11-251-1/+0
| | | | | * remove DefaultValueBuilderFactory * remove PackedMixedTensorBuilderFactory and all packed mixed code.
* Merge pull request #15412 from ↵Arne H Juul2020-11-211-29/+80
|\ | | | | | | | | vespa-engine/havardpe/improved-benchmarking-fairness Havardpe/improved benchmarking fairness
| * use same loop_cnt when benchmarking if possibleHåvard Pettersen2020-11-201-6/+47
| |
| * each EvalOp gets its own stash, for more fairnessHåvard Pettersen2020-11-201-23/+33
| |
* | track CellType moveArne Juul2020-11-201-1/+1
|/
* combine dimensions and split reduce operationsHåvard Pettersen2020-11-191-1/+30
|
* move "keep as-is" optimizersArne Juul2020-11-121-0/+6
| | | | | * from eval/tensor/dense to eval/instruction * minimal changes to track move in this commit
* benchmark some forms of join with numberArne Juul2020-11-101-0/+18
|
* untangle factory-based optimization pipeline from DefaultTensorEngineHåvard Pettersen2020-11-031-22/+23
|
* drop BM of PackedMixedTensorBuilderFactoryArne Juul2020-10-261-3/+0
|
* use a run-time flag instead of conditional compilationArne Juul2020-10-264-27/+15
|
* partial duplicate of micro-bemchmarkArne Juul2020-10-254-7/+27
| | | | | | | * use conditional compilation to let the installed vespa-tensor-instructions-benchmark run its measurements only for the two most important implementations (old prod and new prod).
* Merge pull request #15012 from ↵Arne H Juul2020-10-221-20/+36
|\ | | | | | | | | vespa-engine/havardpe/better-dense-plan-for-generic-reduce improve generic dense reduce with more robust cell ordering
| * improve generic dense reduce with more robust cell orderingHåvard Pettersen2020-10-221-20/+36
| | | | | | | | | | | | | | - unroll reduce all loop - extend benchmark with optimized instructions (to catch slow ones) - drop SimpleTensorEngine from benchmark (always very slow) - tweak API on templated aggregators
* | rename to vespa-tensor-instructions-benchmark and install itArne Juul2020-10-222-2/+4
|/
* added mixed -> partial mixed peek casesHåvard Pettersen2020-10-161-0/+2
|
* added tensor peek benchmarkHåvard Pettersen2020-10-161-1/+94
|
* added tensor lambda benchmarkHåvard Pettersen2020-10-161-1/+58
|
* added encode/decode benchmarkHåvard Pettersen2020-10-161-0/+71
|
* added tensor create benchmarkHåvard Pettersen2020-10-161-41/+86
|
* extend map benchmark with number caseHåvard Pettersen2020-10-161-1/+6
|
* benchmark GenericMap alsoArne Juul2020-10-131-0/+45
|
* allow interpreted function to use new generic operationsHåvard Pettersen2020-10-121-61/+23
| | | | update conformance test to also test with SimpleValue and FastValue
* Merge pull request #14769 from vespa-engine/arnej/fix-concat-collapsingArne H Juul2020-10-081-0/+71
|\ | | | | Arnej/fix concat collapsing
| * benchmark concatArne Juul2020-10-081-0/+71
| |
* | fast value to enable inlined sparse operationsHåvard Pettersen2020-10-071-4/+7
|/ | | | | | | use full overlap sparse join as initial test of full inlining. also improve simple sparse map performance by pre-calculating string hashes and using hash values for equality checks.
* benchmark mergeHåvard Pettersen2020-10-021-4/+68
|
* generic reduceHåvard Pettersen2020-10-021-19/+168
|
* Implement new Value API in SparseTensorArne Juul2020-10-011-0/+6
| | | | | | | | | | * new Address -> index mapping in SparseTensorIndex * extra indirection in SparseTensor * rename old "apply" utilities -> join * make a celltype-templated SparseTensorT and its Builder * add large vector sparse multiply benchmark * get rid of temporary SparseTensorValue * handle templated DirectSparseTensorBuilder in searchlib
* improve benchmark reportHåvard Pettersen2020-09-291-14/+93
|
* Merge pull request #14592 from vespa-engine/arnej/new-sparse-tensor-value-2Arne H Juul2020-09-281-3/+3
|\ | | | | Arnej/new sparse tensor value 2
| * less asserts and parametersArne Juul2020-09-281-4/+0
| |
| * just hold std::vector<T> inside SparseTensorValueArne Juul2020-09-281-0/+4
| |
| * benchmark with new "adaptive" factoryArne Juul2020-09-281-3/+3
| |
* | more generic tensor spec generationHåvard Pettersen2020-09-281-21/+22
|/
* instruction benchmarkHåvard Pettersen2020-09-243-0/+329