OpenMPSupport.rst source code [clang_source_code/docs/OpenMPSupport.rst]

1	.. raw:: html
2
3	<style type="text/css">
4	.none { background-color: #FFCCCC }
5	.partial { background-color: #FFFF99 }
6	.good { background-color: #CCFF99 }
7	</style>
8
9	.. role:: none
10	.. role:: partial
11	.. role:: good
12
13	.. contents::
14	:local:
15
16	==================
17	OpenMP Support
18	==================
19
20	Clang supports the following OpenMP 5.0 features
21
22	* The `reduction`-based clauses in the `task` and `target`-based directives.
23
24	* Support relational-op != (not-equal) as one of the canonical forms of random
25	access iterator.
26
27	* Support for mapping of the lambdas in target regions.
28
29	* Parsing/sema analysis for the requires directive.
30
31	* Nested declare target directives.
32
33	* Make the `this` pointer implicitly mapped as `map(this[:1])`.
34
35	* The `close` map-type-modifier.
36
37	Clang fully supports OpenMP 4.5. Clang supports offloading to X86_64, AArch64,
38	PPC64[LE] and has `basic support for Cuda devices`_.
39
40	* #pragma omp declare simd: :partial:`Partial`. We support parsing/semantic
41	analysis + generation of special attributes for X86 target, but still
42	missing the LLVM pass for vectorization.
43
44	In addition, the LLVM OpenMP runtime `libomp` supports the OpenMP Tools
45	Interface (OMPT) on x86, x86_64, AArch64, and PPC64 on Linux, Windows, and macOS.
46
47	General improvements
48	--------------------
49	- New collapse clause scheme to avoid expensive remainder operations.
50	Compute loop index variables after collapsing a loop nest via the
51	collapse clause by replacing the expensive remainder operation with
52	multiplications and additions.
53
54	- The default schedules for the `distribute` and `for` constructs in a
55	parallel region and in SPMD mode have changed to ensure coalesced
56	accesses. For the `distribute` construct, a static schedule is used
57	with a chunk size equal to the number of threads per team (default
58	value of threads or as specified by the `thread_limit` clause if
59	present). For the `for` construct, the schedule is static with chunk
60	size of one.
61
62	- Simplified SPMD code generation for `distribute parallel for` when
63	the new default schedules are applicable.
64
65	.. _basic support for Cuda devices:
66
67	Cuda devices support
68	====================
69
70	Directives execution modes
71	--------------------------
72
73	Clang code generation for target regions supports two modes: the SPMD and
74	non-SPMD modes. Clang chooses one of these two modes automatically based on the
75	way directives and clauses on those directives are used. The SPMD mode uses a
76	simplified set of runtime functions thus increasing performance at the cost of
77	supporting some OpenMP features. The non-SPMD mode is the most generic mode and
78	supports all currently available OpenMP features. The compiler will always
79	attempt to use the SPMD mode wherever possible. SPMD mode will not be used if:
80
81	- The target region contains an `if()` clause that refers to a `parallel`
82	directive.
83
84	- The target region contains a `parallel` directive with a `num_threads()`
85	clause.
86
87	- The target region contains user code (other than OpenMP-specific
88	directives) in between the `target` and the `parallel` directives.
89
90	Data-sharing modes
91	------------------
92
93	Clang supports two data-sharing models for Cuda devices: `Generic` and `Cuda`
94	modes. The default mode is `Generic`. `Cuda` mode can give an additional
95	performance and can be activated using the `-fopenmp-cuda-mode` flag. In
96	`Generic` mode all local variables that can be shared in the parallel regions
97	are stored in the global memory. In `Cuda` mode local variables are not shared
98	between the threads and it is user responsibility to share the required data
99	between the threads in the parallel regions.
100
101	Collapsed loop nest counter
102	---------------------------
103
104	When using the collapse clause on a loop nest the default behavior is to
105	automatically extend the representation of the loop counter to 64 bits for
106	the cases where the sizes of the collapsed loops are not known at compile
107	time. To prevent this conservative choice and use at most 32 bits,
108	compile your program with the `-fopenmp-optimistic-collapse`.
109
110
111	Features not supported or with limited support for Cuda devices
112	---------------------------------------------------------------
113
114	- Cancellation constructs are not supported.
115
116	- Doacross loop nest is not supported.
117
118	- User-defined reductions are supported only for trivial types.
119
120	- Nested parallelism: inner parallel regions are executed sequentially.
121
122	- Static linking of libraries containing device code is not supported yet.
123
124	- Automatic translation of math functions in target regions to device-specific
125	math functions is not implemented yet.
126
127	- Debug information for OpenMP target regions is supported, but sometimes it may
128	be required to manually specify the address class of the inspected variables.
129	In some cases the local variables are actually allocated in the global memory,
130	but the debug info may be not aware of it.
131
132

Clang Project