forked from OpenMP/Examples
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Examples_async_target.tex
54 lines (41 loc) · 3.06 KB
/
Examples_async_target.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
\pagebreak
\chapter{Asynchronous Execution of a \code{target} Region Using Tasks}
\label{chap:async_target}
The following example shows how the \code{task} and \code{target} constructs
are used to execute multiple \code{target} regions asynchronously. The task that
encounters the \code{task} construct generates an explicit task that contains
a \code{target} region. The thread executing the explicit task encounters a task
scheduling point while waiting for the execution of the \code{target} region
to complete, allowing the thread to switch back to the execution of the encountering
task or one of the previously generated explicit tasks.
\cexample{async_target}{1c}
The Fortran version has an interface block that contains the \code{declare} \code{target}.
An identical statement exists in the function declaration (not shown here).
\fexample{async_target}{1f}
The following example shows how the \code{task} and \code{target} constructs
are used to execute multiple \code{target} regions asynchronously. The task dependence
ensures that the storage is allocated and initialized on the device before it is
accessed.
\cexample{async_target}{2c}
The Fortran example below is similar to the C version above. Instead of pointers, though, it uses
the convenience of Fortran allocatable arrays on the device. An allocatable array has the
same behavior in a \code{map} clause as a C pointer, in this case.
If there is no shape specified for an allocatable array in a \code{map} clause, only the array descriptor
(also called a dope vector) is mapped. That is, device space is created for the descriptor, and it
is initially populated with host values. In this case, the \plc{v1} and \plc{v2} arrays will be in a
non-associated state on the device. When space for \plc{v1} and \plc{v2} is allocated on the device
the addresses to the space will be included in their descriptors.
At the end of the first \code{target} region, the descriptor (of an unshaped specification of an allocatable
array in a \code{map} clause) is returned with the raw device address of the allocated space.
The content of the array is not returned. In the example the data in arrays \plc{v1} and \plc{v2}
are not returned. In the second \code{target} directive, the \plc{v1} and \plc{v2} descriptors are
re-created on the device with the descriptive information; and references to the
vectors point to the correct local storage, of the space that was not freed in the first \code{target}
directive. At the end of the second \code{target} region, the data in array \plc{p} is copied back
to the host since \plc{p} is not an allocatable array.
A \code{depend} clause is used in the \code{task} directive to provide a wait at the beginning of the second
\code{target} region, to insure that there is no race condition with \plc{v1} and \plc{v2} in the two tasks.
It would be noncompliant to use \plc{v1} and/or \plc{v2} in lieu of \plc{N} in the \code{depend} clauses,
because the use of non-allocated allocatable arrays as list items in the first \code{depend} clause would
lead to unspecified behavior.
\fexample{async_target}{2f}