Commit message (Collapse) | Author | Age | Files | Lines | |
---|---|---|---|---|---|
* | Merge pull request #15536 from ↵ | Håvard Pettersen | 2020-11-30 | 3 | -12/+20 |
|\ | | | | | | | | | vespa-engine/havardpe/improve-reference-evaluation Havardpe/improve reference evaluation | ||||
| * | simplify reference lambda | Håvard Pettersen | 2020-11-30 | 1 | -8/+4 |
| | | |||||
| * | add some tests and normalize everything | Håvard Pettersen | 2020-11-30 | 2 | -4/+16 |
| | | |||||
* | | Merge pull request #15512 from vespa-engine/arnej/renumber-peek-children | Arne H Juul | 2020-11-30 | 3 | -8/+12 |
|\ \ | |/ |/| | renumber Peek children | ||||
| * | renumber Peek children | Arne Juul | 2020-11-27 | 3 | -8/+12 |
| | | | | | | | | | | | | * in the Peek::Spec, children now start at 1. 0 is reserved for the "input" (peeked child) parameter. This corresponds to child numbering in the tensor function. | ||||
* | | normalize peek result to fix out-of-bounds dense subspaces | Håvard Pettersen | 2020-11-27 | 3 | -5/+2 |
| | | |||||
* | | reference evaluation | Håvard Pettersen | 2020-11-27 | 16 | -15/+693 |
| | | |||||
* | | check with SimpleValueBuilderFactory and always do type resolving | Arne Juul | 2020-11-27 | 1 | -9/+8 |
|/ | |||||
* | test dense optimizers with FastValue instead of DefaultTensorEngine | Arne Juul | 2020-11-27 | 12 | -54/+45 |
| | |||||
* | test with FastValue instead of DefaultTensorEngine | Arne Juul | 2020-11-27 | 1 | -2/+2 |
| | |||||
* | test with FastValue instead of DefaultTensorEngine | Arne Juul | 2020-11-27 | 6 | -116/+0 |
| | |||||
* | test with FastValue instead of DefaultTensorEngine | Arne Juul | 2020-11-27 | 3 | -12/+7 |
| | |||||
* | put aggregator before dimensions in reference reduce() | Arne Juul | 2020-11-26 | 4 | -20/+20 |
| | |||||
* | split some test cases | Arne Juul | 2020-11-26 | 1 | -28/+79 |
| | |||||
* | make several "test cases" | Arne Juul | 2020-11-26 | 1 | -13/+20 |
| | |||||
* | no testing with float cells here | Arne Juul | 2020-11-26 | 1 | -60/+56 |
| | |||||
* | test ReferenceOperations implementations | Arne Juul | 2020-11-26 | 3 | -0/+450 |
| | |||||
* | add normalize() method to tensorspec | Arne Juul | 2020-11-25 | 5 | -22/+93 |
| | |||||
* | Merge pull request #15461 from vespa-engine/arnej/lift-reference-operations | Arne H Juul | 2020-11-25 | 18 | -249/+455 |
|\ | | | | | move reference operations to a common place | ||||
| * | move spec normalization out from reference operation to unit test | Arne Juul | 2020-11-25 | 4 | -18/+20 |
| | | |||||
| * | move Peek and Create specs | Arne Juul | 2020-11-25 | 8 | -48/+70 |
| | | | | | | | | | | | | | | | | | | * definition of generic Spec for Peek and Create now inside their respective tensor functions * rename internal map [spec() -> map()] * add utility make_spec() to the tensor function to create a generic Spec * use Spec from tensor functions in GenericCreate and GenericPeek * use Spec from tensor functions in ReferenceOperations | ||||
| * | move reference operations to a common place | Arne Juul | 2020-11-25 | 11 | -207/+389 |
| | | | | | | | | | | | | | | * lift reference implementations from unit tests * write a reference TensorCreate implementation * use these from unit tests * some minor cosmetic changes with structured bindings | ||||
* | | Merge branch 'master' into arnej/add-simple-streamed-value | Arne H Juul | 2020-11-25 | 42 | -1607/+32 |
|\| | |||||
| * | GC unused code | Arne Juul | 2020-11-25 | 27 | -1586/+2 |
| | | | | | | | | | | * remove DefaultValueBuilderFactory * remove PackedMixedTensorBuilderFactory and all packed mixed code. | ||||
| * | update test name | Arne Juul | 2020-11-24 | 1 | -1/+1 |
| | | |||||
| * | Flip default to FastValueBuilderFactory | Arne Juul | 2020-11-24 | 3 | -8/+8 |
| | | |||||
| * | cblas may need initialized memory | Arne Juul | 2020-11-24 | 4 | -7/+7 |
| | | |||||
| * | use create_uninitialized_array | Arne Juul | 2020-11-23 | 13 | -19/+19 |
| | | |||||
| * | Reduce inlining for FastSparseMap. | Tor Egge | 2020-11-23 | 2 | -0/+9 |
| | | |||||
* | | fixups after review | Arne Juul | 2020-11-25 | 5 | -13/+15 |
| | | |||||
* | | account for indirect memory usage | Arne Juul | 2020-11-24 | 1 | -1/+5 |
| | | |||||
* | | minor fixups: | Arne Juul | 2020-11-24 | 1 | -2/+2 |
| | | | | | | | | | | * no need for nbostream_longlivedbuf * explicitly get largest possible size_t | ||||
* | | more elegant filter; move assert earlier | Arne Juul | 2020-11-24 | 1 | -5/+2 |
| | | |||||
* | | fix arguments (names and explicit) | Arne Juul | 2020-11-24 | 1 | -6/+9 |
| | | |||||
* | | Fixes after review | Arne Juul | 2020-11-24 | 2 | -12/+16 |
| | | | | | | | | | | | | | | | | | | | | | | | | for StreamedValue: * num_ss -> num_subspaces * require num_mapped_dimensions as constructor argument for StreamedValueBuilder: * do not add the required dense subspace here; it must be added by the caller * remember _num_mapped_dimensions * _dsss -> _dense_subspace_size | ||||
* | | remove superfluous additions | Arne Juul | 2020-11-24 | 1 | -34/+0 |
| | | |||||
* | | add documentation comments | Arne Juul | 2020-11-24 | 5 | -0/+30 |
| | | |||||
* | | run conformance test for StreamedValue also | Arne Juul | 2020-11-23 | 1 | -0/+6 |
| | | |||||
* | | add unit test for StreamedValue | Arne Juul | 2020-11-23 | 3 | -0/+146 |
| | | |||||
* | | add CellTypeUtils class with various static constexpr methods | Arne Juul | 2020-11-23 | 2 | -2/+72 |
| | | |||||
* | | add StreamedValue | Arne Juul | 2020-11-23 | 15 | -0/+463 |
|/ | |||||
* | minor optimization of sparse_only_merge | Arne Juul | 2020-11-23 | 3 | -19/+21 |
| | |||||
* | Merge pull request #15410 from ↵ | Arne H Juul | 2020-11-21 | 1 | -4/+15 |
|\ | | | | | | | | | vespa-engine/geirst/mixed-partial-update-remove-test Add partial update remove test with mixed tensor as input. | ||||
| * | Add partial update remove test with mixed tensor as input. | Geir Storli | 2020-11-20 | 1 | -4/+15 |
| | | |||||
* | | Merge pull request #15412 from ↵ | Arne H Juul | 2020-11-21 | 1 | -29/+80 |
|\ \ | | | | | | | | | | | | | vespa-engine/havardpe/improved-benchmarking-fairness Havardpe/improved benchmarking fairness | ||||
| * | | use same loop_cnt when benchmarking if possible | Håvard Pettersen | 2020-11-20 | 1 | -6/+47 |
| | | | |||||
| * | | each EvalOp gets its own stash, for more fairness | Håvard Pettersen | 2020-11-20 | 1 | -23/+33 |
| |/ | |||||
* | | remove backwards compatibility | Arne Juul | 2020-11-20 | 1 | -1/+0 |
| | | |||||
* | | track CellType move | Arne Juul | 2020-11-20 | 22 | -41/+33 |
|/ | |||||
* | Merge pull request #15391 from vespa-engine/arnej/move-celltype-out-1 | Arne H Juul | 2020-11-20 | 4 | -29/+45 |
|\ | | | | | move CellType up to eval namespace |