B.2  Version 4.5 to 5.0 Differences

• The memory model was extended to distinguish different types of flush operations according to specified flush properties (see Section 1.4.4 on page 59) and to define a happens before order based on synchronizing flush operations (see Section 1.4.5 on page 63).
• Various changes throughout the specification were made to provide initial support of C11, C++11, C++14, C++17 and Fortran 2008 (see Section 1.7 on page 83).
• Fortran 2003 is now fully supported (see Section 1.7 on page 83).
• The requires directive (see Section 2.4 on page 165) was added to support applications that require implementation-specific features.
• The target-offload-var internal control variable (see Section 2.5 on page 171) and the OMP_TARGET_OFFLOAD environment variable (see Section 6.17 on page 1680) were added to support runtime control of the execution of device constructs.
• Control over whether nested parallelism is enabled or disabled was integrated into the max-active-levels-var internal control variable (see Section 2.5.2 on page 189), the default value of which is now implementation defined, unless determined according to the values of the OMP_NUM_THREADS (see Section 6.2 on page 1650) or OMP_PROC_BIND (see Section 6.4 on page 1653) environment variables.
• Support for array shaping (see Section 2.1.4 on page 115) and for array sections with non-unit strides in C and C++ (see Section 2.1.5 on page 118) was added to facilitate specification of discontiguous storage and the target update construct (see Section 2.12.6 on page 497) and the depend clause (see Section 2.17.11 on page 736) were extended to allow the use of shape-operators (see Section 2.1.4 on page 115).
• Iterators (see Section 2.1.6 on page 122) were added to support expressions in a list that expand to multiple expressions.
• The metadirective directive (see Section 2.3.4 on page 154) and declare variant directive (see Section 2.3.5 on page 159) were added to support selection of directive variants and declared function variants at a callsite, respectively, based on compile-time traits of the enclosing context.
• The teams construct (see Section 2.7 on page 236) was extended to support execution on the host device without an enclosing target construct (see Section 2.12.5 on page 478).
• The canonical loop form was defined for Fortran and, for all base languages, extended to permit non-rectangular loop nests (see Section 2.9.1 on page 271).
• The relational-op in the canonical loop form for C/C++ was extended to include != (see Section 2.9.1 on page 271).
• The default loop schedule modifier for worksharing-loop constructs without the static schedule and the ordered clause was changed to nonmonotonic (see Section 2.9.2 on page 288).
• The collapse of associated loops that are imperfectly nested loops was defined for the worksharing-loop (see Section 2.9.2 on page 288), simd (see Section 2.9.3.1 on page 316), taskloop (see Section 2.10.2 on page 395) and distribute (see Section 2.9.4.2 on page 350) constructs.
• The simd construct (see Section 2.9.3.1 on page 316) was extended to accept the if, nontemporal and order(concurrent) clauses and to allow the use of atomic constructs within it.
• The loop construct and the order(concurrent) clause were added to support compiler optimization and parallelization of loops for which iterations may execute in any order, including concurrently (see Section 2.9.5 on page 363).
• The scan directive (see Section 2.9.6 on page 376) and the inscan modifier for the reduction clause (see Section 2.19.5.4 on page 885) were added to support inclusive and exclusive scan computations.
• To support task reductions, the task (see Section 2.10.1 on page 384) and target (see Section 2.12.5 on page 478) constructs were extended to accept the in_reduction clause (see Section 2.19.5.6 on page 897), the taskgroup construct (see Section 2.17.6 on page 648) was extended to accept the task_reduction clause Section 2.19.5.5 on page 896), and the task modifier was added to the reduction clause (see Section 2.19.5.4 on page 885).
• The affinity clause was added to the task construct (see Section 2.10.1 on page 384) to support hints that indicate data affinity of explicit tasks.
• The detach clause for the task construct (see Section 2.10.1 on page 384) and the omp_fulfill_event runtime routine (see Section 3.5.1 on page 1184) were added to support execution of detachable tasks.
• To support taskloop reductions, the taskloop (see Section 2.10.2 on page 395) and taskloop simd (see Section 2.10.3 on page 407) constructs were extended to accept the reduction (see Section 2.19.5.4 on page 885) and in_reduction (see Section 2.19.5.6 on page 897) clauses.
• The taskloop construct (see Section 2.10.2 on page 395) was added to the list of constructs that can be canceled by the cancel construct (see Section 2.18.1 on page 752)).
• To support mutually exclusive inout sets, a mutexinoutset dependence-type was added to the depend clause (see Section 2.10.6 on page 415 and Section 2.17.11 on page 736).
• Predefined memory spaces (see Section 2.11.1 on page 427), predefined memory allocators and allocator traits (see Section 2.11.2 on page 428) and directives, clauses (see Section 2.11 on page 426 and API routines (see Section 3.7 on page 1208) to use them were added to support different kinds of memories.
• The semantics of the use_device_ptr clause for pointer variables was clarified and the use_device_addr clause for using the device address of non-pointer variables inside the target data construct was added (see Section 2.12.2 on page 451).
• To support reverse offload, the ancestor modifier was added to the device clause for target constructs (see Section 2.12.5 on page 478).
• To reduce programmer effort implicit declare target directives for some functions (C, C++, Fortran) and subroutines (Fortran) were added (see Section 2.12.5 on page 478 and Section 2.12.7 on page 513).
• The target update construct (see Section 2.12.6 on page 497) was modified to allow array sections that specify discontiguous storage.
• The to and from clauses on the target update construct (see Section 2.12.6 on page 497), the depend clause on task generating constructs (see Section 2.17.11 on page 736), and the map clause (see Section 2.19.7.1 on page 939) were extended to allow any lvalue expression as a list item for C/C++.
• Support for nested declare target directives was added (see Section 2.12.7 on page 513).
• New combined constructs master taskloop (see Section 2.13.7 on page 549), parallel master (see Section 2.13.6 on page 546), parallel master taskloop (see Section 2.13.9 on page 554), master taskloop simd (see Section 2.13.8 on page 551), parallel master taskloop simd (see Section 2.13.10 on page 556) were added.
• The depend clause was added to the taskwait construct (see Section 2.17.5 on page 642).
• To support acquire and release semantics with weak memory ordering, the acq_rel, acquire, and release clauses were added to the atomic construct (see Section 2.17.7 on page 652) and flush construct (see Section 2.17.8 on page 677), and the memory ordering semantics of implicit flushes on various constructs and runtime routines were clarified (see Section 2.17.8.1 on page 684).
• The atomic construct was extended with the hint clause (see Section 2.17.7 on page 652).
• The depend clause (see Section 2.17.11 on page 736) was extended to support iterators and to support depend objects that can be created with the new depobj construct.
• Lock hints were renamed to synchronization hints, and the old names were deprecated (see Section 2.17.12 on page 746).
• To support conditional assignment to lastprivate variables, the conditional modifier was added to the lastprivate clause (see Section 2.19.4.5 on page 862).
• The description of the map clause was modified to clarify the mapping order when multiple map-types are specified for a variable or structure members of a variable on the same construct. The close map-type-modifier was added as a hint for the runtime to allocate memory close to the target device (see Section 2.19.7.1 on page 939).
• The capability to map C/C++ pointer variables and to assign the address of device memory that is mapped by an array section to them was added. Support for mapping of Fortran pointer and allocatable variables, including pointer and allocatable components of variables, was added (see Section 2.19.7.1 on page 939).
• The defaultmap clause (see Section 2.19.7.2 on page 978) was extended to allow selecting the data-mapping or data-sharing attributes for any of the scalar, aggregate, pointer or allocatable classes on a per-region basis. Additionally it accepts the none parameter to support the requirement that all variables referenced in the construct must be explicitly mapped or privatized.
• The declare mapper directive was added to support mapping of data types with direct and indirect members (see Section 2.19.7.3 on page 982).
• The omp_set_nested (see Section 3.2.10 on page 1052) and omp_get_nested (see Section 3.2.11 on page 1055) routines and the OMP_NESTED environment variable (see Section 6.9 on page 1665) were deprecated.
• The omp_get_supported_active_levels routine was added to query the number of active levels of parallelism supported by the implementation (see Section 3.2.15 on page 1067).
• Runtime routines omp_set_affinity_format (see Section 3.2.30 on page 1109), omp_get_affinity_format (see Section 3.2.31 on page 1112), omp_set_affinity (see Section 3.2.32 on page 1115), and omp_capture_affinity (see Section 3.2.33 on page 1119) and environment variables OMP_DISPLAY_AFFINITY (see Section 6.13 on page 1671) and OMP_AFFINITY_FORMAT (see Section 6.14 on page 1674) were added to provide OpenMP runtime thread affinity information.
• The omp_get_device_num runtime routine (see Section 3.2.37 on page 1130) was added to support determination of the device on which a thread is executing.
• The omp_pause_resource and omp_pause_resource_all runtime routines were added to allow the runtime to relinquish resources used by OpenMP (see Section 3.2.43 on page 1144 and Section 3.2.44 on page 1150).
• Support for a first-party tool interface (see Section 4 on page 1229) was added.
• Support for a third-party tool interface (see Section 5 on page 1485) was added.
• Support for controlling offloading behavior with the OMP_TARGET_OFFLOAD environment variable was added (see Section 6.17 on page 1680).
• Stubs for Runtime Library Routines(previously Appendix A) were moved to a separate document.
• Interface Declarations (previously Appendix B) were moved to a separate document.