Declaring the scope attributes of variables in an OpenMP parallel region is called scoping. In general, if a variable is scoped as SHARED, all threads share a single copy of the variable. If a variable is scoped as PRIVATE, each thread has its own copy of the variable. OpenMP has a rich data environment. In addition to SHARED and PRIVATE, the scope of a variable can also be declared FIRSTPRIVATE, LASTPRIVATE, REDUCTION, or THREADPRIVATE.
OpenMP requires the user to declare the scope of each variable used in a parallel region. This is a tedious and error-prone process and many find this to be the hardest part of using OpenMP to parallelize programs.
The Sun Studio C, C++, and Fortran 95 compilers provide an automatic scoping feature. The compilers analyze the execution and synchronization pattern of a parallel region and determine automatically what the scope of a variable should be, based on a set of scoping rules.
The autoscoping data scope clause is a Sun extension to the OpenMP specification. A user can specify a variable to be autoscoped by using one of the following two clauses.
The __auto clause on a parallel construct directs the compiler to automatically determine the scope of the named variables in the construct. (Note the two underscores before auto.)
The __auto clause can appear on a PARALLEL, PARALLEL DO/for, PARALLEL SECTIONS, or on a Fortran 95 PARALLEL WORKSHARE directive.
If a variable is specified on the __auto clause, then it cannot be specified in any other data scope clause.
The default(__auto) clause on a parallel construct directs the compiler to automatically determine the scope of all variables referenced in the construct that are not explicitly scoped in any data scope clause.
The default(__auto) clause can appear on a PARALLEL, PARALLEL DO/for, PARALLEL SECTIONS, or on a Fortran 95 PARALLEL WORKSHARE directive.
These rules do not apply to variables whose scopes are predetermined by the OpenMP specification, such as loop iteration variables of worksharing DO or FOR loops. Refer to OpenMP 3.0 Specification (section 220.127.116.11, page 78) for a complete listing of variables whose scopes are predetermined.
S1: If the use of the variable in the parallel region is free of data race conditions for the threads in the team executing the region, then the variable is scoped SHARED.
S2: If in each thread executing the parallel region, the variable is always written before being read by the same thread, then the variable is scoped PRIVATE. The variable is scoped as LASTPRIVATE if it can be scoped PRIVATE and is read before it is written after the parallel region, and the construct is either a PARALLEL DO or a PARALLEL SECTIONS.
S3: If the variable is used in a reduction operation that can be recognized by the compiler, then the variable is scoped REDUCTION with that particular operation type.
A1: If the use of the array in the parallel region is free of data race conditions for the threads in the team executing the region, then the array is scoped as SHARED.
When autoscoping a variable that does not have predetermined scope, the compiler checks the use of the variable against the above rules S1–S3 in the given order if it is a scalar, and against the above rule A1 if it is an array. If a rule matches, the compiler will scope the variable according to the matching rule. If a rule does not match, the compiler tries the next rule. If the compiler is unable to find a match, the compiler gives up attempting to determine the scope of that variable and it is scoped SHARED and the binding parallel region is serialized as if an IF (.FALSE.) or if(0) clause were specified.
There are two reasons why autoscoping fails. One is that the use of the variable does not match any of the rules. The other is that the source code is too complex for the compiler to do a sufficient analysis. Function calls, complicated array subscripts, memory aliasing, and user-implemented synchronizations are some typical causes. (See 6.5 Known Limitations of the Current Implementation.)
Autoscoping in C and C++ applies only to basic data types: integer, floating point, and pointer. If a user specifies a structure variable or class variable to be autoscoped, the compiler will scope the variable as shared and the enclosing parallel region will be executed by a single thread.
The compiler will produce an inline commentary when compiled with the -g debug option. This generated commentary can be viewed with the er_src command, as shown below. (The er_src command is provided as part of the Sun Studio software; for more information, see the er_src(1) man page or the Sun Studio Performance Analyzer manual.)
A good place to start is to compile with the -xvpara option. A warning message will be printed out if autoscoping fails, as shown below.
%cat t.f INTEGER X(100), Y(100), I, T C$OMP PARALLEL DO DEFAULT(__AUTO) DO I=1, 100 T = Y(I) CALL FOO(X) X(I) = T*T END DO C$OMP END PARALLEL DO END %f95 -xopenmp -xO3 -vpara -c t.f "t.f", line 2: Warning: parallel region will be executed by a single thread because the autoscoping of following variables failed - x
Compile with -vpara with f95, -xvpara with cc. (This option has not yet been implemented in CC.)
%cat t.f INTEGER X(100), Y(100), I, T C$OMP PARALLEL DO DEFAULT(__AUTO) DO I=1, 100 T = Y(I) X(I) = T*T END DO C$OMP END PARALLEL DO END %f95 -xopenmp -xO3 -g -c t.f %er_src t.o Source file: ./t.f Object file: ./ot.o Load Object: ./t.o 1. INTEGER X(100), Y(100), I, T Source OpenMP region below has tag R1 Variables autoscoped as SHARED in R1: x, y Variables autoscoped as PRIVATE in R1: t, i Private variables in R1: i, t Shared variables in R1: y, x 2. C$OMP PARALLEL DO DEFAULT(__AUTO) <Function: _$d1A2.MAIN_> Source loop below has tag L1 L1 parallelized by explicit user directive L1 parallel loop-body code placed in function _$d1A2.MAIN_ along with 0 inner loops Copy in M-function of loop below has tag L2 L2 scheduled with steady-state cycle count = 3 L2 unrolled 4 times L2 has 0 loads, 0 stores, 2 prefetches, 0 FPadds, 0 FPmuls, and 0 FPdivs per iteration L2 has 1 int-loads, 1 int-stores, 4 alu-ops, 1 muls, 0 int-divs and 1 shifts per iteration 3. DO I=1, 100 4. T = Y(I) 5. X(I) = T*T 6. END DO 7. C$OMP END PARALLEL DO 8. END
Next, a more complicated example to illustrate how the autoscoping rules work.
1. REAL FUNCTION FOO (N, X, Y) 2. INTEGER N, I 3. REAL X(*), Y(*) 4. REAL W, MM, M 5. 6. W = 0.0 7. 8. C$OMP PARALLEL DEFAULT(__AUTO) 9. 10. C$OMP SINGLE 11. M = 0.0 12. C$OMP END SINGLE 13. 14. MM = 0.0 15. 16. C$OMP DO 17. DO I = 1, N 18. T = X(I) 19. Y(I) = T 20. IF (MM .GT. T) THEN 21. W = W + T 22. MM = T 23. END IF 24. END DO 25. C$OMP END DO 26. 27. C$OMP CRITICAL 28. IF ( MM .GT. M ) THEN 29. M = MM 30. END IF 31. C$OMP END CRITICAL 32. 33. C$OMP END PARALLEL 34. 35. FOO = W - M 36. 37. RETURN 38. END
The function FOO() contains a parallel region, which contains a SINGLE construct, a work-sharing DO construct and a CRITICAL construct. If we ignore all the OpenMP parallel constructs, what the code in the parallel region does is:
Copy the value in array X to array Y
Find the maximum positive value in X, and store it in M
Accumulate the value of some elements of X into variable W.
Let's see how the compiler uses the autoscoping rules in 6.2 Autoscoping Rulesto find the appropriate scopes for the variables in the parallel region.
The following variables are used in the parallel region, I, N, MM, T, W, M, X, and Y. The compiler will determine the following.
Scalar I is the loop iteration variable of the work-sharing DO loop. According to the OpenMP specification, the scope of I is predetermined to be PRIVATE.
Scalar N is only read in the parallel region and therefore will not cause data race, so it is scoped as SHARED following rule S1.
Any thread executing the parallel region will execute statement 14, which sets the value of scalar MM to 0.0. This write will cause data race, so rule S1 does not apply. The write happens before any read of MM in the same thread, so MM is scoped as PRIVATE according to rule S2.
Similarly, scalar T is scoped as PRIVATE.
Scalar W is read and then written at statement 21, so rules S1 and S2 do not apply. The addition operation is both associative and communicative, therefore, W is scoped as REDUCTION(+) according to rule S3.
Scalar M is written in statement 11 which is inside a SINGLE construct. The implicit barrier at the end of the SINGLE construct ensures that the write in statement 11 will not happen concurrently with either the read in statement 28 or the write in statement 29, and the latter two will not happen at the same time because both are inside the same CRITICAL construct. No two threads can access M at the same time. Therefore, the writes and reads of M in the parallel region do not cause a data race, and, following rule S1, M is scoped SHARED.
Array X is only read and not written in the region, so it is scoped as SHARED by rule A1.
The writes to array Y is distributed among the threads, and no two threads will write to the same elements of Y. As there is no data race, Y is scoped SHARED according to rule A1.
Only OpenMP directives are recognized and used in the analysis. Calls to OpenMP runtime routines are not recognized. For example, if a program uses OMP_SET_LOCK() and OMP_UNSET_LOCK() to implement a critical section, the compiler is not able to detect the existence of the critical section. Use CRITICAL and END CRITICAL directives if possible.
Only synchronizations specified by using OpenMP synchronization directives, such as BARRIER and MASTER, are recognized and used in the analysis. User-implemented synchronizations, such as busy-waiting, are not recognized.
Autoscoping is not supported when compiling with -xopenmp=noopt.