cuflow/sched.h: SMP Parallelization
[SMP Work Scheduling]

Data Structures
struct	cuflow_exeq_entry
struct	cuflow_exeq
Defines
#define	CUFLOW_EXEQ_SIZE   8
#define	CUFLOW_EXEQ_MASK   (CUFLOW_EXEQ_SIZE - 1)
#define	CUFLOW_PROFILE_SCHED   0
#define	CUFLOW_CALLS_BETWEEN_SCHED   1
#define	cuflow_exeqpri_succ(pri)   ((cuflow_exeqpri_t)((pri) + 1))
Typedefs
typedef struct cuflow_exeq *	cuflow_exeq_t
typedef struct cuflow_exeq_entry *	cuflow_exeq_entry_t
Enumerations
enum	cuflow_exeqpri_t { cuflow_exeqpri_begin = 0, cuflow_exeqpri_normal = 0, cuflow_exeqpri_background = 1, cuflow_exeqpri_end }
Functions
cuflow_exeqpri_t	cuflow_sched_priority (void)
cuflow_exeqpri_t	cuflow_sched_change_priority (cuflow_exeqpri_t priority)
cuflow_exeq_t	cuflow_sched_exeq (void)
void	cuflow_sched_call_on (cu_clop0(f, void), AO_t *cdisj, cuflow_exeq_t exeq)
void	cuflow_sched_call_at (cu_clop0(f, void), AO_t *cdisj, cuflow_exeqpri_t pri)
void	cuflow_sched_call (cu_clop0(f, void), AO_t *cdisj)
void	cuflow_sched_call_sub1_on (cu_clop0(f, void), AO_t *cdisj, cuflow_exeq_t exeq)
void	cuflow_sched_call_sub1_at (cu_clop0(f, void), AO_t *cdisj, cuflow_exeqpri_t pri)
void	cuflow_sched_call_sub1 (cu_clop0(f, void), AO_t *cdisj)
cu_bool_t	cuflow_exeqpri_prioreq (cuflow_exeqpri_t pri0, cuflow_exeqpri_t pri1)
cu_bool_t	cuflow_exeqpri_prior (cuflow_exeqpri_t pri0, cuflow_exeqpri_t pri1)

---

Detailed Description

This provides an efficient mechanism for sharing work across multiple CPUs or cores on an SMP system. Before use, start some worker threads with cuflow_workers_spawn_at_least or cuflow_workers_spawn. To parallelize a task, initialise a local cuflow_cdisj_t guard to zero and request the subtasks with cuflow_sched_call passing a reference to guard. Then use cuflow_cdisj_wait_while to wait for the subtasks to finish.

---

Enumeration Type Documentation

enum cuflow_exeqpri_t

A type to hold the priority of scheduled work.

---

Function Documentation

cu_bool_t cuflow_exeqpri_prior	(	cuflow_exeqpri_t	pri0,
		cuflow_exeqpri_t	pri1
	)

True iff pri0 is a higher priority than pri1.

cu_bool_t cuflow_exeqpri_prioreq	(	cuflow_exeqpri_t	pri0,
		cuflow_exeqpri_t	pri1
	)

True iff pri0 is at least as high a priority as pri1.

void cuflow_sched_call	(	cu_clop0(f, void)	,
		AO_t *	cdisj
	)

Same as cuflow_sched_call_on with the current thread-local execution queue passed as the last argument.

void cuflow_sched_call_at	(	cu_clop0(f, void)	,
		AO_t *	cdisj,
		cuflow_exeqpri_t	pri
	)

Same as cuflow_sched_call with the current priority temporarily set to pri.

void cuflow_sched_call_on	(	cu_clop0(f, void)	,
		AO_t *	cdisj,
		cuflow_exeq_t	exeq
	)

If exeq queue is full, calls f and exits, else increments cdisj and schedules f for later execution, possibly by another thread. In the latter case, *cdisj is decremented after f has been called.

This function does not alter the priority of the current thread. It's main purpose is as an optimisation of successive calls to cuflow_sched_call by letting the client fetch the thread-local exeq with cuflow_sched_exeq once and pass it.

void cuflow_sched_call_sub1	(	cu_clop0(f, void)	,
		AO_t *	cdisj
	)

Same as cuflow_sched_call_sub1_on with the current thread-local execution queue passed as the last argument.

void cuflow_sched_call_sub1_at	(	cu_clop0(f, void)	,
		AO_t *	cdisj,
		cuflow_exeqpri_t	pri
	)

Same as cuflow_sched_call_sub1 with the current priority temporarily set to pri.

void cuflow_sched_call_sub1_on	(	cu_clop0(f, void)	,
		AO_t *	cdisj,
		cuflow_exeq_t	exeq
	)

Same as cuflow_sched_call, except that after f has been called, cdisj will be decremented by one. Use this variant if you initialise cdisj in advance to the number of pending calls.

cuflow_exeqpri_t cuflow_sched_change_priority

(

cuflow_exeqpri_t

priority

)

Set the priority at which to run work scheduled by the current thread, and return the old priority.

cuflow_exeq_t cuflow_sched_exeq

(

void

)

The current current execution queue on which work will be scheduled.

cuflow_exeqpri_t cuflow_sched_priority

(

void

)

The priority at which to run work scheduled by the current thread.