Draft: Fortran Support

Testing using level 2

Disqualified Tests:

• Passing NULL/MPI_REQUEST_NULL causes compile errors in Fortran, disqualifying InvalidParam tests that do this
  • Disqualifies 17 tests (lvl1) / 136 (lvl2)
• A lot of tests are currently removed due to datatypes, i.e., MPI C types that do not exist / are not implemented yet for Fortran
  • No effect on lvl1, as only MPI_INT and MPI_DOUBLE are used / 501 disqualified (lvl2)

Unfixable, but not disqualified:

• InvalidParam-Status-mpi_recv-002: MPI_STATUSES_IGNORE rejected by compiler, different rank than expected
  • Error class does not exist for Fortran. Disqualifies 1 test (lvl1) / 16 (lvl2)

MPI Issues:

• https://github.com/open-mpi/ompi/issues/9534 - Causes issues with Precv_Init / Psend_init. Need recent OpenMPI version to avoid this
• https://github.com/open-mpi/ompi/pull/13308 - Bug in OpenMPI-git causes compilation failure as well

Types are translated from C to Fortran where applicable, though currently only for MPI_INT -> MPI_INTEGER, MPI_DOUBLE -> MPI_DOUBLE_PRECISION, MPI_C_BOOL -> MPI_LOGICAL as most Fortran MPI types are optional. Tests are formatted using fprettify (available from pip, added to Pipfile).

Common Issues encountered:

• MPI does not differentiate between bool and int, but does between integer and logical.
  • Workaround: Use bool where applicable in C code, then translate to logical
• Different integer size expectations cause compilation failures (no automatic extension)
  • Workaround: During parsing of apis.json, save the integer width, and perform explicit widening if needed.
• C usage of address-of-integer as a one-dimensional array does not work in Fortran
  • Workaround: Replace uses thereof with proper arrays
• Array Indexing starts at 1
  • Workaround: Implement explicit indexing to match C

MPI-BugBench bugs (since remembering to write them down):

• Deadlock in CallOrdering-mpi_improbe-mpi_send-001, wrong parameter used as flag
  • Note: Deadlock also happens in InvalidParam-flag-mpi_iprobe-002, but intended? (not_allocated flag is not waited on, but the "normal" one)

scripts/Infrastructure/MPICall.py

@@ -6,21 +6,12 @@ import re
 from typing_extensions import override
 
 from Infrastructure.Instruction import Instruction
-from Infrastructure.Variables import ERROR_MARKER_COMMENT_BEGIN, ERROR_MARKER_COMMENT_END, ERROR_MARKER_COMMENT_BEGIN_FORT, ERROR_MARKER_COMMENT_END_FORT
+from Infrastructure.Variables import ERROR_MARKER_COMMENT_BEGIN, ERROR_MARKER_COMMENT_END, ERROR_MARKER_COMMENT_BEGIN_FORT, ERROR_MARKER_COMMENT_END_FORT, arg_c_to_fort
 import Infrastructure.Variables as infvars
 
 call_template_fort = "call @FUNC@(@ARGS@)"
 call_template_c    = "@FUNC@(@ARGS@);"
 
-arg_c_to_fort = {
-    "MPI_INT": "MPI_INTEGER",
-    "MPI_DOUBLE": "MPI_DOUBLE_PRECISION",
-    "sizeof(int)": "int_size",
-    "sizeof(signed int)": "int_size",
-    "sizeof(double)": "double_size",
-    "true": ".TRUE.",
-    "false": ".FALSE.",
-}
 
 def get_arg_str(args: typing.OrderedDict[str, str], bigint: typing.List[str] = []):
     s = ""
@@ -30,6 +21,9 @@ def get_arg_str(args: typing.OrderedDict[str, str], bigint: typing.List[str] = [
             # NULL() and MPI_REQUEST_NULL (as a literal, not variable) are both caught by the compiler
             # Tests are invalid, and would not compile
             if v == "NULL" or v == "MPI_REQUEST_NULL": v = "! FORTRAN_INCOMPATIBLE"
+            # Using a C datatype with no (implemented) fortran equivalent, also invalid
+            from Infrastructure.TemplateFactory import predefined_types
+            if v in predefined_types and v not in arg_c_to_fort: v = "! FORTRAN_INCOMPATIBLE"
             for c_arg in arg_c_to_fort.keys():
                 if c_arg in str(v):
                     v = v.replace(c_arg, arg_c_to_fort[c_arg])