xref: /llvm-project/openmp/docs/openacc/OpenMPExtensions.rst (revision 4a368136693ba9c4e827702e9d390280c3d5e7ac)
1OpenMP Extensions for OpenACC
2=============================
3
4OpenACC provides some functionality that OpenMP does not.  In some
5cases, Clang supports OpenMP extensions to provide similar
6functionality, taking advantage of the runtime implementation already
7required for OpenACC.  This section documents those extensions.
8
9By default, Clang recognizes these extensions.  The command-line
10option ``-fno-openmp-extensions`` can be specified to disable all
11OpenMP extensions, including those described in this section.
12
13.. _ompx-motivation:
14
15Motivation
16----------
17
18There are multiple benefits to exposing OpenACC functionality as LLVM
19OpenMP extensions:
20
21* OpenMP applications can take advantage of the additional
22  functionality.
23* As LLVM's implementation of these extensions matures, it can serve
24  as a basis for including these extensions in the OpenMP standard.
25* Source-to-source translation from certain OpenACC features to OpenMP
26  is otherwise impossible.
27* Runtime tests can be written in terms of OpenMP instead of OpenACC
28  or low-level runtime calls.
29* More generally, there is a clean separation of concerns between
30  OpenACC and OpenMP development in LLVM.  That is, LLVM's OpenMP
31  developers can discuss, modify, and debug LLVM's extended OpenMP
32  implementation and test suite without directly considering OpenACC's
33  language and execution model, which are handled by LLVM's OpenACC
34  developers.
35
36.. _ompx-hold:
37
38``ompx_hold`` Map Type Modifier
39-------------------------------
40
41.. _ompx-holdExample:
42
43Example
44^^^^^^^
45
46.. code-block:: c++
47
48  #pragma omp target data map(ompx_hold, tofrom: x) // holds onto mapping of x throughout region
49  {
50    foo(); // might have map(delete: x)
51    #pragma omp target map(present, alloc: x) // x is guaranteed to be present
52    printf("%d\n", x);
53  }
54
55The ``ompx_hold`` map type modifier above specifies that the ``target
56data`` directive holds onto the mapping for ``x`` throughout the
57associated region regardless of any ``target exit data`` directives
58executed during the call to ``foo``.  Thus, the presence assertion for
59``x`` at the enclosed ``target`` construct cannot fail.
60
61.. _ompx-holdBehavior:
62
63Behavior
64^^^^^^^^
65
66* Stated more generally, the ``ompx_hold`` map type modifier specifies
67  that the associated data is not unmapped until the end of the
68  construct.  As usual, the standard OpenMP reference count for the
69  data must also reach zero before the data is unmapped.
70* If ``ompx_hold`` is specified for the same data on lexically or
71  dynamically enclosed constructs, there is no additional effect as
72  the data mapping is already held throughout their regions.
73* The ``ompx_hold`` map type modifier is permitted to appear only on
74  ``target`` constructs (and associated combined constructs) and
75  ``target data`` constructs.  It is not permitted to appear on
76  ``target enter data`` or ``target exit data`` directives because
77  there is no associated statement, so it is not meaningful to hold
78  onto a mapping until the end of the directive.
79* The runtime reports an error if ``omp_target_disassociate_ptr`` is
80  called for a mapping for which the ``ompx_hold`` map type modifier
81  is in effect.
82* Like the ``present`` map type modifier, the ``ompx_hold`` map type
83  modifier applies to an entire struct if it's specified for any
84  member of that struct even if other ``map`` clauses on the same
85  directive specify other members without the ``ompx_hold`` map type
86  modifier.
87* ``ompx_hold`` support is not yet provided for ``defaultmap``.
88
89Implementation
90^^^^^^^^^^^^^^
91
92* LLVM uses the term *dynamic reference count* for the standard OpenMP
93  reference count for host/device data mappings.
94* The ``ompx_hold`` map type modifier selects an alternate reference
95  count, called the *hold reference count*.
96* A mapping is removed only once both its reference counts reach zero.
97* Because ``ompx_hold`` can appear only constructs, increments and
98  decrements of the hold reference count are guaranteed to be
99  balanced, so it is impossible to decrement it below zero.
100* The dynamic reference count is used wherever ``ompx_hold`` is not
101  specified (and possibly cannot be specified).  Decrementing the
102  dynamic reference count has no effect if it is already zero.
103* The runtime determines that the ``ompx_hold`` map type modifier is
104  *in effect* (see :ref:`Behavior <ompx-holdBehavior>` above) when the
105  hold reference count is greater than zero.
106
107Relationship with OpenACC
108^^^^^^^^^^^^^^^^^^^^^^^^^
109
110OpenACC specifies two reference counts for tracking host/device data
111mappings.  Which reference count is used to implement an OpenACC
112directive is determined by the nature of that directive, either
113dynamic or structured:
114
115* The *dynamic reference count* is always used for ``enter data`` and
116  ``exit data`` directives and corresponding OpenACC routines.
117* The *structured reference count* is always used for ``data`` and
118  compute constructs, which are similar to OpenMP's ``target data``
119  and ``target`` constructs.
120
121Contrast with OpenMP, where the dynamic reference count is always used
122unless the application developer specifies an alternate behavior via
123our map type modifier extension.  We chose the name *hold* for that
124map type modifier because, as demonstrated in the above :ref:`example
125<ompx-holdExample>`, *hold* concisely identifies the desired behavior
126from the application developer's perspective without referencing the
127implementation of that behavior.
128
129The hold reference count is otherwise modeled after OpenACC's
130structured reference count.  For example, calling ``acc_unmap_data``,
131which is similar to ``omp_target_disassociate_ptr``, is an error when
132the structured reference count is not zero.
133
134While Flang and Clang obviously must implement the syntax and
135semantics for selecting OpenACC reference counts differently than for
136selecting OpenMP reference counts, the implementation is the same at
137the runtime level.  That is, OpenACC's dynamic reference count is
138OpenMP's dynamic reference count, and OpenACC's structured reference
139count is our OpenMP hold reference count extension.
140
141.. _atomicWithinTeams:
142
143``atomic`` Strictly Nested Within ``teams``
144-------------------------------------------
145
146Example
147^^^^^^^
148
149OpenMP 5.2, sec. 10.2 "teams Construct", p. 232, L9-12 restricts what
150regions can be strictly nested within a ``teams`` region.  As an
151extension, Clang relaxes that restriction in the case of the
152``atomic`` construct so that, for example, the following case is
153permitted:
154
155.. code-block:: c++
156
157  #pragma omp target teams map(tofrom:x)
158  #pragma omp atomic update
159  x++;
160
161Relationship with OpenACC
162^^^^^^^^^^^^^^^^^^^^^^^^^
163
164This extension is important when translating OpenACC to OpenMP because
165OpenACC does not have the same restriction for its corresponding
166constructs.  For example, the following is conforming OpenACC:
167
168.. code-block:: c++
169
170  #pragma acc parallel copy(x)
171  #pragma acc atomic update
172  x++;
173