OpenMP Reference Sheet for C/C++ <A,B,C such that total iterations known at start of loop>

for(A=C;A<B;A++) {
<your code here>

Constructs <force ordered execution of part of the code. A=C will be guaranteed to execute
<parallelize a for loop by breaking apart iterations into chunks> before A=C+1>
#pragma omp parallel for [shared(vars), private(vars), firstprivate(vars), #pragma omp ordered {
lastprivate(vars), default(shared|none), reduction(op:vars), copyin(vars), if(expr), <your code here>
ordered, schedule(type[,chunkSize])] }
<A,B,C such that total iterations known at start of loop> }
for(A=C;A<B;A++) {
<your code here> <parallelized sections of code with each section operating in one thread>
#pragma omp sections [private(vars), firstprivate(vars), lastprivate(vars),
<force ordered execution of part of the code. A=C will be guaranteed to execute reduction(op:vars), nowait] {
before A=C+1> #pragma omp section {
#pragma omp ordered { <your code here>
<your code here> }
} #pragma omp section {
} <your code here>
}
<parallelized sections of code with each section operating in one thread> ....
#pragma omp parallel sections [shared(vars), private(vars), firstprivate(vars), }
lastprivate(vars), default(shared|none), reduction(op:vars), copyin(vars), if(expr)] {
#pragma omp section { <only one thread will execute the following. NOT always by the master thread>
<your code here> #pragma omp single {
} <your code here (only executed once)>
#pragma omp section { }
<your code here> }
}
....
} Directives
shared(vars) <share the same variables between all the threads>
<grand parallelization region with optional work-sharing constructs defining more private(vars) <each thread gets a private copy of variables. Note that other than the
specific splitting of work and variables amongst threads. You may use work-sharing master thread, which uses the original, these variables are not initialized to
constructs without a grand parallelization region, but it will have no effect (sometimes anything.>
useful if you are making OpenMP'able functions but want to leave the creation of threads firstprivate(vars) <like private, but the variables do get copies of their master thread
to the user of those functions)> values>
#pragma omp parallel [shared(vars), private(vars), firstprivate(vars), lastprivate(vars), lastprivate(vars) <copy back the last iteration (in a for loop) or the last section (in a
default(private|shared|none), reduction(op:vars), copyin(vars), if(expr)] { sections) variables to the master thread copy (so it will persist even after the
<the work-sharing constructs below can appear in any order, are optional, and can parallelization ends)>
be used multiple times. Note that no new threads will be created by the constructs. default(private|shared|none) <set the default behavior of variables in the parallelization
They reuse the ones created by the above parallel construct.> construct. shared is the default setting, so only the private and none setting have
effects. none forces the user to specify the behavior of variables. Note that even with
<your code here (will be executed by all threads)> shared, the iterator variable in for loops still is private by necessity >
reduction(op:vars) <vars are treated as private and the specified operation(op, which
<parallelize a for loop by breaking apart iterations into chunks> can be +,*,-,&,|,&,&&,||) is performed using the private copies in each thread. The
#pragma omp for [private(vars), firstprivate(vars), lastprivate(vars), master thread copy (which will persist) is updated with the final value.>
reduction(op:vars), ordered, schedule(type[,chunkSize]), nowait]
copyin(vars) <used to perform the copying of threadprivate vars to the other threads.
Similar to firstprivate for private vars.>
if(expr) <parallelization will only occur if expr evaluates to true.>
schedule(type [,chunkSize]) <thread scheduling model>
Function Based Locking < nest versions allow recursive locking>
type chunkSize void omp_init_[nest_]lock(omp_lock_t*) <make a generic mutex lock>
static number of iterations per thread pre-assigned at beginning of loop void omp_destroy_[nest_]lock(omp_lock_t*) <destroy a generic mutex lock>
(typical default is number of processors) void omp_set_[nest_]lock(omp_lock_t*) <block until mutex lock obtained>
void omp_unset_[nest_]lock(omp_lock_t*) <unlock the mutex lock>
dynamic number of iterations to allocate to a thread when available (typical int omp_test_[nest_]lock(omp_lock_t*) <is lock currently locked by somebody>
default is 1)
guided highly dependent on specific implementation of OpenMP

nowait <remove the implicit barrier which forces all threads to finish before continuation Settings and Control
in the construct> int omp_get_num_threads() <returns the number of threads used for the parallel
region in which the function was called>
int omp_get_thread_num() <get the unique thread number used to handle this
iteration/section of a parallel construct. You may break up algorithms into parts
based on this number.>
Synchronization/Locking Constructs <May be used almost anywhere, but will
int omp_in_parallel() <are you in a parallel construct>
only have effects within parallelization constructs.>
int omp_get_max_threads() <get number of threads OpenMP can make>
int omp_get_num_procs() <get number of processors on this system>
<only the master thread will execute the following. Sometimes useful for special handling
int omp_get_dynamic() <is dynamic scheduling allowed>
of variables which will persist after the parallelization.>
int omp_get_nested() <is nested parallelism allowed>
#pragma omp master {
double omp_get_wtime() <returns time (in seconds) of the system clock>
<your code here (only executed once and by the master thread).
double omp_get_wtick() <number of seconds between ticks on the system clock>
}
void omp_set_num_threads(int) <set number of threads OpenMP can make>
void omp_set_dynamic(int) <allow dynamic scheduling (note this does not make
<mutex lock the region. name allows the creation of unique mutex locks.>
dynamic scheduling the default)>
#pragma omp critical [(name)] {
void omp_set_nested(int) <allow nested parallelism; Parallel constructs within other
<your code here (only one thread allowed in at a time)>
parallel constructs can make new threads (note this tends to be unimplemented
}
in many OpenMP implementations)>
<force all threads to complete their operations before continuing>
<env vars- implementation dependent, but here are some common ones>
#pragma omp barrier
OMP_NUM_THREADS "number" <maximum number of threads to use>
OMP_SCHEDULE "type,chunkSize" <default #pragma omp schedule settings>
<like critical, but only works for simple operations and structures contained in one line of
code>
#pragma omp atomic
<simple code operation, ex. a += 3; Typical supported operations are ++,--,+,*,-
,/,&,^,<<,>>,| on primitive data types> Legend
vars is a comma separated list of variables
<force a register flush of the variables so all threads see the same memory> [optional parameters and directives]
#pragma omp flush[(vars)] <descriptions, comments, suggestions>
.... above directive can be used multiple times
<applies the private clause to the vars of any future parallelize constructs encountered (a For mistakes, suggestions, and comments please email e_berta@plutospin.com
convenience routine)>
#pragma omp threadprivate(vars)