aboutsummaryrefslogtreecommitdiffstats
path: root/eval
Commit message (Collapse)AuthorAgeFilesLines
* allow serializing dense tensor viewsHåvard Pettersen2018-02-059-13/+23
| | | | | needed for fall-back to reference implementation using on-the-fly generated dense tensors that are not of the 'DenseTensor' class.
* co-locate optimization with implementationHåvard Pettersen2018-02-029-136/+113
|
* go back to allocating result on working stashArne Juul2018-02-022-12/+10
|
* minor cleanupHåvard Pettersen2018-02-011-11/+3
|
* pre-alloc optimized xw product output vectorHåvard Pettersen2018-02-012-18/+27
|
* remove tensor function eval testsHåvard Pettersen2018-01-311-14/+0
| | | | covered by normal interpreted function tests
* use interpreted function to evaluate tensor functionHåvard Pettersen2018-01-311-2/+6
|
* use interpreted function to evaluate tensor functionHåvard Pettersen2018-01-311-2/+3
|
* remove tensor function eval from cross-language conformance testHåvard Pettersen2018-01-311-21/+0
|
* use interpreted function to evaluate tensor functionHåvard Pettersen2018-01-311-2/+6
|
* added test constructor to make interpreted function from tensor functionHåvard Pettersen2018-01-312-0/+13
|
* remove retained eval from conformance testsHåvard Pettersen2018-01-311-85/+0
| | | | (as it is now equivalent to interpreted evaluation)
* remove recursive eval from optimized tensor function operationsHåvard Pettersen2018-01-314-27/+0
|
* remove recursive eval from tensor functionHåvard Pettersen2018-01-312-77/+0
|
* allow any intermediate result as input for inner productsHåvard Pettersen2018-01-311-13/+11
|
* optimized XW product now has childrenHåvard Pettersen2018-01-305-75/+99
| | | | also remove instruction trampoline
* optimized dot product now has childrenHåvard Pettersen2018-01-306-26/+35
| | | | also remove instruction trampoline
* special case for parameter fetchingHåvard Pettersen2018-01-303-25/+27
| | | | state replace -> pop_push and pop_pop_push
* use optimized+compiled tensor function in interpreted functionHåvard Pettersen2018-01-2617-450/+81
| | | | rename 'compile' to 'optimize' in tensor engine
* compile tensor functionHåvard Pettersen2018-01-267-1/+266
|
* Merge pull request #4763 from vespa-engine/bratseth/support-binary-tensor-filesJon Bratseth2018-01-243-2/+13
|\ | | | | Support binary tensor files with ending .tbf
| * Remove unnecessary includesJon Bratseth2018-01-241-2/+0
| |
| * Spaces onlyJon Bratseth2018-01-241-8/+8
| |
| * Use memorymapped fileJon Bratseth2018-01-241-23/+19
| |
| * Support binary tensor files with ending .tbfJon Bratseth2018-01-243-17/+34
| |
* | improve type resolving for 'if'Håvard Pettersen2018-01-198-18/+89
|/ | | | also verify inferred types against actual result in conformance test
* run cross-language tensor conformance tests using tensor functionsHåvard Pettersen2018-01-1813-54/+59
| | | | also pass tensor engine to tensor function eval function
* run all tests from test_spec as tensor functionsHåvard Pettersen2018-01-181-16/+32
|
* added initial version of make_tensor_functionHåvard Pettersen2018-01-184-0/+325
|
* test new tensor IR nodesHåvard Pettersen2018-01-171-11/+126
|
* added 'const_value' and 'if_node' to tensor IRHåvard Pettersen2018-01-162-12/+83
|
* add 'concat' and 'rename' tensor operationsHåvard Pettersen2018-01-162-41/+137
| | | | also add intermediate Leaf/Op1/Op2 classes
* tweak tensor IR semantics and add more encapsulationHåvard Pettersen2018-01-168-140/+170
|
* move more param stuff to lazy_params.hHåvard Pettersen2018-01-169-36/+41
|
* Use lazy parameters in tensor functions.Håvard Pettersen2018-01-1617-79/+67
|
* move LazyParams to separate fileHåvard Pettersen2018-01-166-12/+44
|
* added test for push_children (tensor IR nodes)Håvard Pettersen2018-01-151-0/+24
|
* tensor function compilation based on bottom-up in-place optimizationHåvard Pettersen2018-01-154-12/+75
|
* separate optimize and compile conceptsHåvard Pettersen2018-01-1510-77/+85
| | | | | | | | optimize: tensor function -> tensor function compile: node tree -> tensor function node trees now contain tensor functions to support direct recursive mixed-mode evaluation.
* Merge pull request #4612 from vespa-engine/balder/some-misc-old-cleanupHenning Baldersheim2018-01-112-17/+13
|\ | | | | Use c++11 for loops and alignment.
| * Avoid using temporaries.Henning Baldersheim2018-01-102-17/+13
| |
* | Merge pull request #4595 from vespa-engine/balder/track-common-cell-indexHenning Baldersheim2018-01-113-66/+93
|\ \ | |/ |/| Balder/track common cell index
| * Put the address back in the combiner.Henning Baldersheim2018-01-103-25/+24
| |
| * Add comment explaining a bit about the right address.Henning Baldersheim2018-01-101-0/+2
| |
| * Move the right address context on the outside.Henning Baldersheim2018-01-103-75/+68
| |
| * Use the same AddressContext for the combinedAddress and compute the ↵Henning Baldersheim2018-01-103-30/+45
| | | | | | | | combinedIdx inline too.
| * Hide the calculation inside the address context.Henning Baldersheim2018-01-091-2/+3
| |
| * Make and address context to keep code together.Henning Baldersheim2018-01-092-30/+47
| |
| * Rename to indicate it refers to right side.Henning Baldersheim2018-01-092-10/+10
| |
* | Revert earlier reorder.Henning Baldersheim2018-01-101-2/+2
| |