To propagate CPU-specific settings for floating-point computations to tasks executed by the task scheduler, you can use one of the following two methods:
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.
The task_group_context class has a method to capture the current floating-point settings.
By 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.
For 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:
task_group_context ctx( task_group_context::isolated,
task_group_context::default_traits | task_group_context::fp_settings );
Or call the capture_fp_settings method:
You 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. It is possible to execute the parallel algorithms with different floating-point settings captured to separate contexts, even at the same time.
Floating-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. Otherwise, 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.
In 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. If none of the outer-level contexts capture floating-point settings, the settings captured during task arena initialization are used.
It guarantees that:
Floating-point settings are applied to all tasks executed within a task arena, if they are captured:
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.
The guarantees above apply only to the following conditions:
Otherwise, the stated guarantees are not valid and the behavior related to floating-point settings is undefined.