New Features after OpenMP 2.5

Transcription

1

2 New Features after OpenMP 2.5 2

3 Outline OpenMP Specifications Version 3.0 Task Parallelism Improvements to nested and loop parallelism Additional new Features Version New Features Version 4.0 simd Construct and Reduction Clause Additional new Features Accelerator and Error Model Version Plans 3

4 OpenMP Specifications Version 3.0 released in May New task level parallelism - Improvements to loop and nested parallelism - Additional Clauses, runtime functions and environment variables Version 3.1 released in July Additional Clauses - Improvements to task parallelism - Initial support for thread binding 4

5 OpenMP 4.0 RC 1 in November SIMD directives - Extended support for tasking and affinity control - Initial support for Fortran Initial support for user-defined reductions OpenMP 4.0 RC 2 in March Additional support for tasking, simd, Fortran Initial accelerator support - Initial error model support 5

6 Version Task Parallelism New feature in OpenMP Tasks: Work units executed by the encountering thread or deferred for execution by any other thread. Tasks are composed of - Code - Data - ICVs 6

7 Two activities: packaging and execution Task Construct C/C++: #pragma omp task [clauses] { } Fortran:!$omp task [clauses]!$omp end task Clauses: if, untied, data clauses 7

8 Data Clauses: Implicit rules apply Otherwise C: default(shared none), private(list), firstprivate(list), shared(list) Nesting: Can be nested - Inside parallel regions - Inside other tasks - Inside work-sharings Fortran: default(private firstprivate shared none ), private(list), firstprivate(list), shared(list) 8

9 Task Synchronisation: taskwait: #pragma omp taskwait!$omp taskwait Task behaviour Barrier: #pragma omp barrier!$omp barrier Implicit barrier #pragma omp parallel num_threads(n) { #pragma omp task function_a(); #pragma omp barrier #pragma omp single { #pragma omp task Fibonacci Sequence: fib(0)=0 fib(1)=1 fib(n)=fib(n-1)+fib(n-2), n>1 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, } function_b(); } 9

10 int main(){ int n=30; omp_set_dynamic(0); omp_set_num_threads(4); #pragma omp parallel shared(n) { #pragma omp single printf("fib(%d)=%d\n", n, fib(n)); } } int fib(int n) { int i, j; if(n<2) return n; else{ #pragma omp task shared(i) firstprivate(n) i=fib(n-1); #pragma omp task shared(j) firstprivate(n) }} j=fib(n-2); #pragma omp taskwait return i+j; 10

11 Version Nested parallelism Recall: Parallel regions can be nested, but support for this is implementation dependent Better support for nested parallelism in v3.0 New library routines, environment variables, multiple internal control variables 11

12 Control maximum number of active parallel regions: OMP_MAX_ACTIVE_LEVELS omp_set_max_active_levels() omp_get_max_active_levels() Control maximum number of OpenMP threads: OMP_THREAD_LIMIT omp_get_thread_limit() To obtain information about nested parallelism: omp_get_level(): How many nested parallel regions at this point? 12

13 omp_get_active_level(): How many active (with 2 or more threads) regions? omp_get_ancestor_thread_num(level): Which thread-id was my ancestor? omp_get_team_size(level): How many threads there are at a previous regions? Multiple ICVs: Allows omp_set_num_threads() inside a parallel region #pragma omp parallel num_threads(3) omp_set_num_threads(omp_get_thread_num()+2); #pragma omp parallel foo(); 13

14 Version Loop parallelism Recall: The iterations are distributed over team threads. schedule(static): #pragma omp for nowait schedule(static) for(i=0;i<n;i++) a(i)= #pragma omp for schedule(static) for(i=0;i<n;i++) a(i)= 14

15 schedule(runtime): omp_set_schedule() AUTO schedule: #pragma omp parallel for schedule(auto) for(i=0;i<n;i++) omp_get_schedule() COLLAPSE clause: #pragma omp parallel for collapse(2) for(i=0;i<n;i++) for(j=0;j<m;j++) 15

16 Version Other Features Additional Environment Variables: OMP_STACK_SIZE size [B K M G]: control of children therad s stack size OMP_WAIT_POLICY [ACTIVE PASSIVE]: control of thread s idle behaviour Minor fixes and clarifications to Version 2.5 See OpenMP Specifications for Version 3.0 in openmp.org 16

17 Version New Features New atomics support capture and write functionality Modifications to data environment: intent(in), const-qualified Initial support for thread binding: OMP_PROC_BIND Extensions to OpenMP tasking model - taskyield Construct: C/C++: #pragma omp taskyield Fortran:!$omp taskyield 17

18 - final and mergeable Clauses: Undeferred task: a task for which execution is not deferred with respect to its generating task region Included task: an undeferred task that is sequentially included in generating task region Merged task: a task that has the same data environment as that of its generating task region. (mergeable) Final task: a task that makes all its child tasks become final and included tasks. (final(expression)) - omp_in_final() 18

19 Version simd construct Transform the loop into a simd loop C/C++: #pragma omp simd [clauses] for loops Fortran:!$omp simd [clauses] do loops!$omp end simd Clauses: safelen(length), linear(list), aligned(list), private(list), lastprivate(list), reduction(op:list), collapse(n) declare simd construct: - #pragma omp declare simd [clauses] -!$omp declare simd [clauses] - Clauses: simdlen(length), linear(list), aligned(list), uniform(list), reduction(op:list), inbranch, notinbranch 19

20 #pragma omp declare simd float min (float a, float b) { return a<b? a : b; } loop simd construct: - #pragma omp for simd [clauses] -!$omp do simd [clauses]!$omp end do simd [nowait] void minex (float *a, float *b, float *c, float *d) { #pragma omp parallel for simd for (i = 0; i < N; i++) d[i] = min (a[i], b[i]); } 20

21 Version reduction clause Define custom reductions C/C++: #pragma omp declare reduction(reduction-identifier : typename-list : combiner) [initializer-clause] Fortran:!$omp declare reduction(reductionidentifier : type-list : combiner) [initializer-clause] #pragma omp declare reduction (merge : std::vector<int> : omp_out.insert(omp_out.end(), omp_in.begin(), omp_in.end())) void schedule (std::vector<int> &v, std::vector<int> &filtered) { #pragma omp parallel for reduction (merge : filtered) for (std:vector<int>::iterator it = v.begin(); it < v.end(); it++) if ( filter(*it)) filtered.push_back(*it); } 21

22 Version Additional Features Task Synchronisation: taskgroup: #pragma omp taskgroup!$omp taskgroup!$omp end taskgroup Affinity Support: New Clause to the parallel construct: proc_bind(master close spread) Increased choices for OMP_PROC_BIND Added OMP_PLACES environment variable New RTLs: omp_proc_bind_t omp_get_proc_bind(void) 22

23 Version Accelerator Model target construct: C/C++: #pragma omp target [clauses] Fortran:!$omp target [clauses]!$omp end target Clauses: if, device, map device clause: device(expression) map clause: map ([map-type:] list) (Types: alloc, to, from, tofrom) #pragma omp target device(acc0) map(a, B) #pragma omp parallel for reduction(+:sum) for(int i=0; i<n; i++){ sum+=a[i]*b[i]; } 23

24 target data construct: C/C++: #pragma omp target data [clauses] Fortran:!$omp target data [clauses]!$omp end target data #pragma omp target data map(a[0:rows][0:cols]) target update construct: C/C++: #pragma omp target update [clauses] Fortran:!$omp target update [clauses] #pragma omp target update to(recv_e, recv_w) #pragma omp target update from(send_e, send_w) Clauses: to, from, device, if 24

25 declare target construct: C/C++: #pragma omp declare target function or declaration Fortran:!$omp declare target (subroutine or function) #pragma omp target declare Void P(restrict float Q[][cols], const int i, const int k) {return Q[i][k]*Q[i][k];} New runtime functions: omp_set_default_device() omp_get_default_device() omp_get_num_devices() New environment variable: OMP_DEFAULT_DEVICE 25

26 Version Error Model Cancel construct: C/C++: #pragma omp cancel [clauses] Clauses: parallel, sections, for/do, taskgroup, if Cancellationpoint construct: C/C++: #pragma omp cancellation point [clauses] Clauses: parallel, sections, for/do, taskgroup OMP_CANCELLATION Fortran:!$omp cancel [clauses] Fortran:!$omp cancellation point [clauses] 26

27 Version Plans Support for memory affinity Refinements to accelerator support Additional task/thread synchronization mechanisms Completing extension of OpenMP to Fortran

28 References nmp%20bof%20sc12%20v2.pdf 6. AlexDuran-Transparents-12Mar09.pdf 7. ew_sun.pdf

29 Thank you! Created by Buket Benek Gursoy from ICHEC, Ireland 29