`MPI_Abi_get_fortran_booleans` always returns `MPI_INFO_NULL`

[eschnett]

I am building the pull request #13280 and am using the new MPI ABI. I find that the function MPI_Abi_get_fortran_booleans always returns MPI_INFO_NULL.

I believe the problem lies in the function ompi_abi_get_fortran_booleans, which compares e.g. the constants OMPI_DATATYPE_MPI_LOGICAL1 and OMPI_DATATYPE_MPI_INT8_T in a switch statement. These constants never agree.

The following patch makes things work for me:

diff --git a/ompi/mpi/c/ompi_abi_fortran.c b/ompi/mpi/c/ompi_abi_fortran.c
index da139e7..eb502eb 100644
--- a/ompi/mpi/c/ompi_abi_fortran.c
+++ b/ompi/mpi/c/ompi_abi_fortran.c
@@ -224,96 +224,18 @@ int ompi_abi_get_fortran_booleans(int logical_size, void *logical_true, void *lo
     bool unavailable = false;
     bool use_int8_t = false, use_int16_t = false, use_int32_t = false, use_int64_t = false;
     switch (logical_size) {
-    case OMPI_SIZEOF_FORTRAN_LOGICAL1:
-        switch (OMPI_DATATYPE_MPI_LOGICAL1) {
-        case OMPI_DATATYPE_MPI_INT8_T:
-            use_int8_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT16_T:
-            use_int16_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT32_T:
-            use_int32_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT64_T:
-            use_int64_t = true;
-            break;
-        default:
-            unavailable = true;
-            break;
-         }
-    case OMPI_SIZEOF_FORTRAN_LOGICAL2:
-        switch (OMPI_DATATYPE_MPI_LOGICAL2) {
-        case OMPI_DATATYPE_MPI_INT8_T:
-            use_int8_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT16_T:
-            use_int16_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT32_T:
-            use_int32_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT64_T:
-            use_int64_t = true;
-            break;
-        default:
-            unavailable = true;
-            break;
-        }
-    case OMPI_SIZEOF_FORTRAN_LOGICAL4:
-        switch (OMPI_DATATYPE_MPI_LOGICAL4) {
-        case OMPI_DATATYPE_MPI_INT8_T:
-            use_int8_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT16_T:
-            use_int16_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT32_T:
-            use_int32_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT64_T:
-            use_int64_t = true;
-            break;
-        default:
-            unavailable = true;
-            break;
-        }
-    case OMPI_SIZEOF_FORTRAN_LOGICAL8:
-        switch (OMPI_DATATYPE_MPI_LOGICAL8) {
-        case OMPI_DATATYPE_MPI_INT8_T:
-            use_int8_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT16_T:
-            use_int16_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT32_T:
-            use_int32_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT64_T:
-            use_int64_t = true;
-            break;
-        default:
-            unavailable = true;
-            break;
-        }
-    case OMPI_SIZEOF_FORTRAN_LOGICAL16:
-        switch (OMPI_DATATYPE_MPI_LOGICAL16) {
-        case OMPI_DATATYPE_MPI_INT8_T:
-            use_int8_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT16_T:
-            use_int16_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT32_T:
-            use_int32_t = true;
-            break;
-        case OMPI_DATATYPE_MPI_INT64_T:
-            use_int64_t = true;
-            break;
-        default:
-            unavailable = true;
-            break;
-        }
+    case 1:
+        use_int8_t = true;
+        break;
+    case 2:
+        use_int16_t = true;
+        break;
+    case 4:
+        use_int32_t = true;
+        break;
+    case 8:
+        use_int64_t = true;
+        break;
     default:
         unavailable = true;
     }

[hppritcha]

i'm a little confused. Do you mean if you've set the default error handler to errors return that a call to MPI_Abi_get_fortran_booleans is returning MPI_INFO_NULL or that is_set is returning 0 for particular values of the logical_size argument?

Also are you configuring Open MPI without fortran support?

[eschnett]

I am configuring OpenMPI with Fortran support enabled. I do not want to use its Fortran interfaces (mpi.f90 etc.), but I want support for MPI_INTEGER etc. to be present.

As you suspected, I misspoke in my description. The function MPI_Abi_get_fortran_booleans returns is_set = 0.

[hppritcha]

I think you want to be using MPI_Abi_get_fortran_info and checking for the value of the mpi_integer_size key.
MPI_Abi_get_fortran_boolean returns is_set= 0 if the fortran compiler doesn't support the logical of size logical_size or at least that's the way its supposed to work.

[eschnett]

I think there is a logic error in OpenMPI. If you look at the lines

        switch (OMPI_DATATYPE_MPI_LOGICAL1) {
        case OMPI_DATATYPE_MPI_INT8_T:

then you see that the code compares the two types MPI_LOGICAL1 and MPI_INT8_T. I don't think these types are ever equal. I think what the code wants to do is to find out how many bytes a logical*1 occupies, and write the respective truth value into the provided void* buffer.

[hppritcha]

oh its a bit knarly here. see what's going on in ompi/datatype/ompi_datatype_internal.h

[bosilca]

Unlike OMPI_DATATYPE_MPI_INT8_T which is a predefined type, OMPI_DATATYPE_MPI_LOGICAL1 is a virtual type and it is made equal to a C predefined according to the type size as discovered during configure. The logic is located in ompi/ompi/datatype/ompi_datatype_internal.h, and the set of valid values for OMPI_DATATYPE_MPI_LOGICAL1 is [OMPI_DATATYPE_MPI_LOGICAL, OMPI_DATATYPE_MPI_INT8_T, OMPI_DATATYPE_MPI_INT16_T, OMPI_DATATYPE_MPI_INT32_T, OMPI_DATATYPE_MPI_INT64_T, OMPI_DATATYPE_MPI_UNAVAILABLE].

I think the issue with this code is that it is failing to test for OMPI_DATATYPE_MPI_LOGICAL which should be the default in most cases.

[hppritcha]

The code is looking for the logical size requested in the logical_size arg.

[eschnett]

I believe I misinterpreted the standard. MPI_Abi_get_fortran_booleans should always return information about the Fortran type logical, and not logical*N as I assumed above. The input argument logical_size needs to be the size of the Fortran type logical, in bytes.

[hppritcha]

that is my understanding and the way i hopefully correctly implemented that function.

These MPI ABi fortran functions are primarily intended for developers of 3rd-party MPI fortran bindings like vapaa.

Is that what you are trying to do?

[eschnett]

Yes: https://github.com/eschnett/mpif

[hppritcha]

Okay. In addition to the PR #13280 contents, we will also need to do some additional (much smaller) changes as noted in comment

#13280 (comment)

The MPI_Abi_set_fortran_info/MPI_Abi_set_fortran_booleans in the PR 13280 do not currently actually update the OMPI internal datatype info and would also need to be enhanced to actually do this.

Is your project far enough along that I could use it for testing follow-ons to 13280?

[eschnett]

Yes.

I am using both MPICH and OpenMPI in the CI tests of mpif. I am currently applying a patch to OpenMPI (the patch above, although this is not the correct implementation) to make the call to MPI_Abi_get_fortran_booleans work.

This call is not necessary to make mpif wrok. mpif is not generic with respect to the Fortran integer/logical size. I assume that I know the size ahead of time, and currently also assume that sizeof(MPI_Fint) == 4, which is the default for gfortran and flang.

The mpif CI tests should show how to build and test it.

The largest piece missing in mpif is support for callbacks.