1.. _Floating_Point_Settings:
2
3Floating-point Settings
4=======================
5
6To propagate CPU-specific settings for floating-point computations to tasks executed by the task scheduler, you can use one of the following two methods:
7
8* When a ``task_arena`` or a task scheduler for a given application thread is initialized, they capture the current floating-point settings of the thread.
9* The ``task_group_context`` class has a method to capture the current floating-point settings.
10
11By default, worker threads use floating-point settings obtained during the initialization of a ``task_arena`` or the implicit arena of the application thread. The settings are applied to all computations within that ``task_arena`` or started by that application thread.
12
13
14For better control over floating point behavior, a thread may capture the current settings in a task group context. Do it at context creation with a special flag passed to the constructor:
15
16::
17
18    task_group_context ctx( task_group_context::isolated,
19                        task_group_context::default_traits | task_group_context::fp_settings );
20
21
22Or call the ``capture_fp_settings`` method:
23
24::
25
26     task_group_context ctx;
27    ctx.capture_fp_settings();
28
29
30You can then pass the task group context to most parallel algorithms, including ``flow::graph``, to ensure that all tasks related to this algorithm use the specified floating-point settings.
31It is possible to execute the parallel algorithms with different floating-point settings captured to separate contexts, even at the same time.
32
33Floating-point settings captured to a task group context prevail over the settings captured during task scheduler initialization. It means, if a context is passed to a parallel algorithm, the floating-point settings captured to the context are used.
34Otherwise, if floating-point settings are not captured to the context, or a context is not explicitly specified, the settings captured during the task arena initialization are used.
35
36In a nested call to a parallel algorithm that does not use the context of a task group with explicitly captured floating-point settings, the outer-level settings are used.
37If none of the outer-level contexts capture floating-point settings, the settings captured during task arena initialization are used.
38
39It guarantees that:
40
41* Floating-point settings are applied to all tasks executed within a task arena, if they are captured:
42
43  * To a task group context.
44  * During the arena initialization.
45
46* A call to a oneTBB parallel algorithm does not change the floating-point settings of the calling thread, even if the algorithm uses different settings.
47
48.. note::
49    The guarantees above apply only to the following conditions:
50
51    * A user code inside a task should:
52
53      * Not change the floating-point settings.
54      * Revert any modifications.
55      * Restore previous settings before the end of the task.
56
57    * oneTBB task scheduler observers are not used to set or modify floating point settings.
58
59    Otherwise, the stated guarantees are not valid and the behavior related to floating-point settings is undefined.
60
61