test_target_mapping_alloc relies on missing declare_target

[dreachem]

Describe the bug
The test test_target_mapping_alloc calls test_and_set_verbose in a target construct via the macro OMPVV_TEST_AND_SET_VERBOSE, defined in ompvv.F90. However, the declaration/definition of that procedure does not have a declare_target directive to ensure that compiler generates a device version of it that is callable on a non-host device.

Test that it applies to

Applies to the test_target_mapping_alloc test.

To Reproduce
Failure reproduced when trying to compile this test with CCE. 6.1 will say that support for this is implementation defined. It's not guaranteed to work.

Expected behavior
I expect the test, as written, to potentially not compile or result in a linker error.

Compiler

CCE 20.

Accelerator hardware

Any non-host accelerator will trigger the compile-time error for CCE.

@spophale

[tob2]

I assume that's about 'test_target_mapping_before_alloc.F90', i.e.

OpenMP_VV/tests/5.0/target/test_target_mapping_before_alloc.F90

Lines 50 to 53 in f9cb9f2

	!$omp target map(alloc: scalar, a, test_struct) map(to: scalar, a, &
	!$omp& test_struct) map(tofrom: errors)
	OMPVV_TEST_AND_SET_VERBOSE(errors, scalar .ne. 80 .OR. a(2) .ne. 2 .OR. test_struct%var .ne. 1 .OR. test_struct%b(2) .ne. 2)
	!$omp end target

That file includes ompvv/ompvv.F90, which defines OMPVV_TEST_AND_SET_VERBOSE involving a call to test_and_set_verbose which is a module function in module ompvv_lib.

The code has two potential issues, depending on the compiler + non-host device:

• In verbose mode, it uses I/O ('write' stmt) in the target region
• It assumes that there is an implicit declare target for the function defintion of test_and_set_verbose.

Regarding the latter:

• TR14 has the following wording in "9.4 enter Clause"
  If a procedure name appears in an enter clause in the same compilation unit in which the definition of the procedure occurs then a device-specific version of the procedure is created for all devices to which the directive of the clause applies.
  Whether a device-specific version of a procedure is created is implementation defined when the clause is in a different compilation unit than the definition of the procedure but both are in the same translation unit.

• OpenMP 5.1 has:
  If a procedure appears in a to clause in the same compilation unit in which the definition of the procedure occurs then a device-specific version of the procedure is created.

Thus, in TR14 it is implementation defined, in 5.1 unspecified – however, 5.0 has the following rather vague wording:

If a procedure is referenced in a procedure that appears as a list item in a to clause on a declare target directive then the name of the procedure is treated as if it had appeared in a to clause on a declare target directive.

which can be interpreted as wanted …

The simplest solution is to sacrifice the verbose output and to use:

--- a/tests/5.0/target/test_target_mapping_before_alloc.F90
+++ b/tests/5.0/target/test_target_mapping_before_alloc.F90
@@ -52 +52 @@ CONTAINS
-    OMPVV_TEST_AND_SET_VERBOSE(errors, scalar .ne. 80 .OR. a(2) .ne. 2 .OR. test_struct%var .ne. 1 .OR. test_struct%b(2) .ne. 2)
+    OMPVV_TEST_AND_SET(errors, scalar .ne. 80 .OR. a(2) .ne. 2 .OR. test_struct%var .ne. 1 .OR. test_struct%b(2) .ne. 2)

Still, there is the desire to require the support in the same translation unit, i.e. the compiler might have to implement it in the future. (GCC supports this and – I believe but have not tested it – flag as well; in case of GCC, it is not written into the .mod file (happens to early) but the device version is nonetheless created.)

[seyonglee]

test_and_set_verbose is directly called in a target region, and thus it will follow the following rule:

OpenMP 5.0, Section 2.12.5 target Construct, page 173, line 1 ~ 3:

If a function (C, C++, Fortran) or subroutine (Fortran) is referenced in a target construct then
that function or subroutine is treated as if its name had appeared in a to clause on a
declare target directive.

[dreachem]

test_and_set_verbose is directly called in a target region, and thus it will follow the following rule:

OpenMP 5.0, Section 2.12.5 target Construct, page 173, line 1 ~ 3:

If a function (C, C++, Fortran) or subroutine (Fortran) is referenced in a target construct then
that function or subroutine is treated as if its name had appeared in a to clause on a
declare target directive.

This sentence would only apply within the same compilation unit (program unit) in which the reference occurs. If the definition of the procedure is in a different compilation unit, it still requires a declare target directive so the compiler knows to generate a device version of it.

[seyonglee]

test_and_set_verbose is directly called in a target region, and thus it will follow the following rule:
OpenMP 5.0, Section 2.12.5 target Construct, page 173, line 1 ~ 3:

If a function (C, C++, Fortran) or subroutine (Fortran) is referenced in a target construct then
that function or subroutine is treated as if its name had appeared in a to clause on a
declare target directive.

This sentence would only apply within the same compilation unit (program unit) in which the reference occurs. If the definition of the procedure is in a different compilation unit, it still requires a declare target directive so the compiler knows to generate a device version of it.

As I know, OpenMP 5.0 does not explicitly say that the above sentence (OpenMP 5.0, Section 2.12.5 target Construct, page 173, line 1 ~ 3) is applied only within the same compilation unit (program unit). I think that real problem is that OpenMP 5.0 does not explain how to handle the case where the procedures in the to clause of a declare target directive are not defined in the same compilation unit. As @tob2 mentioned, this behavior is undefined in OpenMP specifications earlier than TR14.

[spophale]

Partially resolved in #900 .