summaryrefslogtreecommitdiffstats
path: root/eval
Commit message (Collapse)AuthorAgeFilesLines
* avoid deadlock when using compile cache with blocking executorsHåvard Pettersen2020-06-143-52/+98
| | | | | | | | - avoid taking cache lock in compile task destructor - avoid holding cache lock while posting compile task - add unit test trying to provoke deadlock - bonus: fix blocking executor implementation (did not work with threads <= task limit)
* Merge pull request #13577 from ↵Henning Baldersheim2020-06-132-5/+6
|\ | | | | | | | | vespa-engine/havardpe/separate-result-locks-for-compile-cache use separate locks for result value propagation
| * use separate locks for result value propagationHåvard Pettersen2020-06-132-5/+6
| | | | | | | | ... to avoid deadlock when using blocking executors
* | Merge pull request #13571 from vespa-engine/arnej/use-typify-invoke-for-matmulHåvard Pettersen2020-06-131-43/+16
|\ \ | | | | | | use typify_invoke to select matmul implementation
| * | replace template magic with if statementHåvard Pettersen2020-06-131-28/+11
| | |
| * | use typify_invoke to select matmul implementationArne Juul2020-06-121-43/+33
| | | | | | | | | | | | | | | | | | * instead of a chain of select functions, use typify_invoke; use partial specialization on a struct to handle the double*double and float*float cases specially.
* | | Merge pull request #13570 from vespa-engine/arnej/use-typify-invoke-for-selectHåvard Pettersen2020-06-138-45/+41
|\ \ \ | | | | | | | | Arnej/use typify invoke for select
| * | | typify_invoke instead of multiple select levelsArne Juul2020-06-131-31/+16
| | | | | | | | | | | | | | | | * follow pattern suggested in matmul PR
| * | | use typify_invoke instead of select_2Arne Juul2020-06-123-6/+10
| | | |
| * | | use typify_invoke instead of select_1Arne Juul2020-06-124-10/+16
| | | |
| * | | use typify_invoke instead of dispatch_0Arne Juul2020-06-121-2/+3
| |/ /
* | / add aggr typifier and use itHåvard Pettersen2020-06-133-30/+33
| |/ |/|
* | even better Inv handling, and testingHåvard Pettersen2020-06-122-2/+2
| |
* | better handling of InvHåvard Pettersen2020-06-122-2/+4
| |
* | inline more operationsHåvard Pettersen2020-06-124-29/+241
|/
* use common apply loopsHåvard Pettersen2020-06-124-33/+28
|
* Use function pointers in inline operations test.Tor Egge2020-06-121-47/+47
|
* use common code for simple map and number joinHåvard Pettersen2020-06-112-86/+22
|
* common code for operation inliningHåvard Pettersen2020-06-1112-126/+391
| | | | | | | | | | - add common code to make selecting the appropriate template function easier (vespa/vespalib/util/typify.h) - enable detection of lambda functions matching all low-level operations. (lookup_op1, lookup_op2) - add typifiers to decide which low-level operations should be inlined (TypifyOp1, TypifyOp2) - integrate into dense_simple_join as a pilot customer
* Revert "Revert "When we pull in a cacheline, we should use it too.""Henning Baldersheim2020-06-081-1/+1
|
* Revert "When we pull in a cacheline, we should use it too."Harald Musum2020-06-081-1/+1
|
* - Stick to void * and byte offsets.Henning Baldersheim2020-06-071-1/+1
| | | | - Correct spelling error.
* added some missing thatsHåvard Pettersen2020-05-291-2/+2
|
* revert unintended changeHåvard Pettersen2020-05-291-1/+1
|
* dense number joinHåvard Pettersen2020-05-298-4/+347
|
* dense tensor lambdaHåvard Pettersen2020-05-287-56/+305
|
* dense simple mapHåvard Pettersen2020-05-2811-180/+231
|
* use index lookup table with shared cacheHåvard Pettersen2020-05-277-22/+307
|
* Move streaming operators to namespace searched by ADL.Tor Egge2020-05-251-0/+4
|
* simple dense joinHåvard Pettersen2020-05-2212-195/+626
|
* let compile cache use shared proton executorHåvard Pettersen2020-05-193-1/+24
|
* dense single reduceHåvard Pettersen2020-05-0712-92/+450
|
* Avoid making copies of container elements.Tor Egge2020-05-041-2/+2
|
* include local file firstHåvard Pettersen2020-05-041-1/+1
|
* multi-matmulHåvard Pettersen2020-05-0411-24/+522
|
* fix PR commentsHåvard Pettersen2020-04-307-13/+8
|
* added float cell range testsHåvard Pettersen2020-04-301-0/+2
|
* lambda peek optimizerHåvard Pettersen2020-04-3017-30/+881
|
* Unwrap reference wrappers to avoid extra indirections viaTor Egge2020-04-231-12/+12
| | | | invalid memory.
* truncate doubles when converting to labels/indexesHåvard Pettersen2020-04-155-9/+9
|
* added skeleton for lambda peek optimizerHåvard Pettersen2020-04-034-0/+34
|
* delay preparing tensor lambda function for executionHåvard Pettersen2020-04-033-13/+25
| | | | | | This will allow implementation-specific tensor lambda optimizations to look at the lambda function and perform appropriate optimizations before it is converted to an interpreted function.
* make tensor engine available when compiling tensor functionsHåvard Pettersen2020-04-0324-55/+67
|
* remove parameter count from interpreted functionsHåvard Pettersen2020-04-035-12/+7
|
* added support for exporting a subset of node typesHåvard Pettersen2020-04-033-5/+78
| | | | | | | This is needed to store type information about tensor lambda inner functions until it is needed; we want to delay making it into an interpreted function until after the actual tensor engine implementation gets a chance to come up with a better optimization.
* Eliminate redundant move in return statement.Tor Egge2020-03-211-1/+1
|
* Merge pull request #12651 from vespa-engine/havardpe/improve-rank-feature-errorsHenning Baldersheim2020-03-202-6/+16
|\ | | | | Havardpe/improve rank feature errors
| * better tensor lambda type errorsHåvard Pettersen2020-03-202-6/+16
| | | | | | | | | | - report actual return type when not double - import type errors from lambda function type resolving
* | ReinlineHenning Baldersheim2020-03-202-16/+9
| |
* | Stick with one way of getting an accelrator.Henning Baldersheim2020-03-201-6/+6
| |