v2.0.0

Interface Changes

Macros

• OPDI_CRITICAL_NAME macro no longer accepts arguments other than name
• new OPDI_CRITICAL_NAME_ARGS macro
• new OPDI_MASKED, OPDI_END_MASKED macros
• new OPDI_SINGLE_COPYPRIVATE, OPDI_SINGLE_COPYPRIVATE_NOWAIT macros

Layer Interfaces

• introduce OmpLogicInterface::beginSkippedParallelRegion, OmpLogicInterface::endSkippedParallelRegion
• refactor OmpLogicInterface::resetTask -> OmpLogicInterface::resetImplicitTask
• changed offset in enums OmpLogicInterface::ScopeEndpoint and OmpLogicInterface::SyncRegionKind
• introduce BackendInterface::getCriticalIdentifier, BackendInterface::getReductionIdentifier, BackendInterface::getOrderedIdentifier
• refactor BackendInterface::getNestedLockIdentifier -> BackendInterface::getNestLockIdentifier
• refactor BackedInterface::getTaskData -> BackendInterface::getImplicitTaskData

Compiler Options

• introduce OPDI_BACKEND_GENERATE_MASKED_EVENTS, OPDI_BACKEND_GENERATE_WORK_EVENTS, OPDI_OMPT_BACKEND_IMPLICIT_TASK_END_SOURCE, OPDI_OMPT_BACKEND_BARRIER_IMPLEMENTATION_BEHAVIOUR, OPDI_OMP_LOGIC_CLEAR_ADJOINTS, OPDI_SYNC_REGION_BARRIER_BEHAVIOUR, OPDI_SYNC_REGION_BARRIER_IMPLICIT_BEHAVIOUR, OPDI_SYNC_REGION_BARRIER_EXPLICIT_BEHAVIOUR, OPDI_SYNC_REGION_BARRIER_IMPLEMENTATION_BEHAVIOUR, OPDI_SYNC_REGION_BARRIER_REVERSE_BEHAVIOUR

Instrumentation

• all functions in OmpLogicInstrumentInterface receive AD data as parameters
• calls to instrumentation occur under the same conditions as AD treatment, like tape activity

Features

• support for custom mutexes
• support for masked construct
• support for broadcasts via single copyprivate
• provide EmptyTool as a default tool layer implementation
• support for disabling OpDiLib at runtime: set opdi::tool to an instance of EmptyTool (preferred) or nullptr
• endpoints of sync regions that result in reverse barriers become configurable
• optionally use the implicit barriers at the end of parallel regions to produce ImplicitTaskEnd events of non-primary threads
• expose mechanism to skip AD treatment of parallel regions
• backend-independent solution for copy operations at the beginning of parallel regions
• optionally clear adjoints as part of tape resets
• the OMPT backend's implementation of opdi_* functions can be included as a default runtime implementation
• syntax checking supports local enablind/disabling via // opdi-syntax-on, // opdi-syntax-off
• additions to syntax definition
• modernize declaration of custom reductions
• faster tests, extended test suite and CI pipeline (see below)
• changes of OmpLogicOutputInstrument in accordance with other changes in OpDiLib

Performance Improvements

• implicit barriers on parallel regions do not produce reverse barriers
• generating events for masked constructs and worksharing constructs becomes optional
• MutexOmpLogic::onMutexReleased no longer acquires an internal mutex
• reductions in the macro backend no longer require additional internal locks, address of parallel data as wait identifier
• ordered constructs in the macro backend no longer share their wait identifier, use address of parallel data as wait identifier

Refactoring and Code Quality

• refactor MutexOmpLogic, more expressive names and reduced boilerplate, comments regarding thread-safety
• refactor ParallelOmpLogic and ImplicitTaskOmpLogic, more expressive names and fewer void* pointers
• transition from master to masked throughout OpDiLib
• remove ProbeScopeStatus, resolve order of ImplicitTaskProbe and ReductionProbe via omp_get_level
• dedicated header file for macros related to thread sanitizer
• refactoring of further smaller aspects throughout OpDiLib
• add warnings, errors, and assertions throughout OpDiLib

Fixes

• ensure that setAdjointAccessMode has no effect when called in the reverse pass
• address compiler warnings
• fix bug in application of syntax checking to a single file
• fix ranges in TestExternalFunctionLocal for small data and/or many threads
• fix memory leak: implicit tasks free allocated tape positions

Tests

Test Suite

• test suite moves from C++11 to C++17
• testing with strict warning levels
• updated image versions
• new tests: TestCriticalNameHint, TestExternalFunctionLogicCalls, TestCustomMutex, TestSingleCopyprivate, TestSingleCopyprivateNowait, TestParallelCopyin, TestParallelFirstprivate, TestParallelFirstprivate2, TestParallelForLastprivate, TestOrderedMultiple, TestOrderedNowait, TestOrderedReduction, TestLock2
• new drivers
  • DriverPrimal for primal computations in the presence of OpDiLib
  • DriverFirstOrderForward for the tapeless forward mode of AD in the presence of OpDiLib
  • DriverFirstOrderReverseVector for vector mode tests
• drivers specify associated tests explicitly
• introduce driver-specific flags
• where possible, prefer OpDiLib's tool interface over direct interaction with the OO AD tool
• faster tests
  • smaller internal data sets
  • default to -O1
  • automatic deduction of appropriate parallelism
  • tests do not use atomics for reversal of serial parts
• instrumentation no longer tied to debug mode
• support for excluding specific tests and drivers from testing
• container file for thread sanitizer tests

CI Pipeline

• test different configurations of OpDiLib
  • source of ImplicitTaskEnd events
  • generation of events for masked constructs, worksharing constructs
• automated syntax checking
• tests with address sanitizer
• tests with thread sanitizer

Miscellaneous

• parallel regions suspend recording on the encountering task's tape
• barriers always produce events for both endpoints
• backend layer becomes responsible for producing events for both reduction-related barriers
• new newsletter address
• Zenodo publication

v1.7.1

Features

• macros that indicate OpDiLib's version

v1.7

Restructure data of parallel regions and implicit tasks

• associate tapes, positions, and adjoint access modes with tasks instead of parallel regions
• extend data of parallel regions by pointers to the parent task and child tasks
• backend support for querying the current task's data
• create task data for initial implicit tasks
• track adjoint access mode entirely via tasks
• support resetting tasks according to a tape position, to recover previous adjoint access modes

Features

• track adjoint access mode of initial implicit task
• support additional sync region and worksharing types
• support use of mutexes before tool initialization (OMPT backend)
• support posting messages that are printed during the reverse pass

Instrument

• implement onSetAdjointAccessMode in OmpLogicOutputInstrument
• adapt OmpLogicInstrumentInterface and OmpLogicOutputInstrument according to the other changes

Tests

• run build tests with enabled output instrument
• tests treat non-empty output on stderr as error (can be disabled)
• add tests' error output files to gitignore
• add assertions to adjoint access control tests
• add TestAdjointAccessControlNested2, TestTaskReset

Fixes

• fix transport of adjoint access mode into and out of parallel regions
• fix using directive in second order test driver
• fix placement of skipParallelHandling decrement

Chore

• update license headers for 2025
• update OpDiLib-related publications in README

v1.6

Features

• Gitlab CI pipeline
• tracking of passive parallel regions
• tsan annotations for mutex synchronization in the reverse pass
• postEvaluate() method in the logic interface

Fixes

• build flags in the test system for producing reference files
• memory leaks
  • in the OPDI_SINGLE macro
  • in tests for external functions
  • clearing the tape pool

Chore

• update links

v1.5.1

• license header update for 2024
• small fix

v1.5

• relicense OpDiLib under LGPLv3

v1.4

• Extend OpDiLib instrumentation.
  • instrumentation of master constructs
  • instrumentation of passive barriers
  • instrumentation of passive parallel regions
  • revise and extend reverse pass instrumentation
• Balance the tape activity calls received by thread 0.

v1.3.2

• update publication
• update tests and examples to CoDiPack 2
• fix OpDiLib finalization order in tests and examples

v1.3.1

• update license header and readme
• improve portability of tests makefile

v1.3

• fix issues regarding reduction combined with nowait
• default adjoint access mode now configurable
• extend tests, test system maintenance
• extend instrumentation, instrumentation maintenance
• various fixes and small changes