HDFGroup
Compilation failure on aarch64 with Arm Compiler for Linux's flang
Using Spack to set up config and compile, the following error occurs: 5 errors found in build log.
It makes no difference if I disable the non-standard float16 - using an arg to cmake of: -DHDF5_ENABLE_NONSTANDARD_FEATURE_FLOAT16:BOOL=OFF
5 errors found in build log:
3047 F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (/tmp/user/spack-stage/s
pack-stage-hdf5-1.14.5-rzhstpbozkrcypkkfn5ntevpg2gwfof7/spack-build-rzhstpb/fortran/static/H5_gen.F90: 604
9)
3048 0 inform, 0 warnings, 1 severes, 0 fatal for h5pget_kind_16
3049 F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (/tmp/user/spack-stage/s
pack-stage-hdf5-1.14.5-rzhstpbozkrcypkkfn5ntevpg2gwfof7/spack-build-rzhstpb/fortran/static/H5_gen.F90: 609
7)
3050 0 inform, 0 warnings, 1 severes, 0 fatal for h5pregister_kind_16
3051 F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (/tmp/user/spack-stage/s
pack-stage-hdf5-1.14.5-rzhstpbozkrcypkkfn5ntevpg2gwfof7/spack-build-rzhstpb/fortran/static/H5_gen.F90: 614
5)
3052 0 inform, 0 warnings, 1 severes, 0 fatal for h5pinsert_kind_16
>> 3053 make[2]: *** [fortran/src/CMakeFiles/hdf5_fortran-static.dir/build.make:325: fortran/src/CMakeFiles/hdf5_f
ortran-static.dir/__/static/H5_gen.F90.o] Error 1
3054 make[2]: Leaving directory '/tmp/user/spack-stage/spack-stage-hdf5-1.14.5-rzhstpbozkrcypkkfn5ntevpg2gw
fof7/spack-build-rzhstpb'
This is using Arm Compiler for Linux - which is LLVM derivative but the NVIDIA/PGI fortran Flang.
As the float16 is disabled, it ought not to be trying to generate it?
I see the same. At configure time I see:
-- Performing Test FORTRAN_CHAR_ALLOC - Success
-- ....NUMBER OF INTEGER KINDS FOUND 4
-- ....REAL KINDS FOUND {4,8,16}
-- ....INTEGER KINDS FOUND {1,2,4,8}
-- ....MAX DECIMAL PRECISION 15
-- ....LOGICAL KINDS FOUND {1,2,4,8}
-- ....FOUND SIZEOF for REAL KINDs {4,8,8}
A simple test tells me that armflang does not support REAL KIND=16, so I'm not sure how REAL KINDS FOUND contains 16. Additionally FOUND SIZEOF for REAL KINDs {4,8,8} is suspicious, since the last two kinds have the same sizeof.
The test seems to happen here: https://github.com/HDFGroup/hdf5/blob/f0cffc90ac0f28818c054db5c467a3cdf2574f58/config/cmake/HDF5UseFortran.cmake#L295
There is a follow-up check here which seems to try and catch the problem:
https://github.com/HDFGroup/hdf5/blob/f0cffc90ac0f28818c054db5c467a3cdf2574f58/config/cmake/HDF5UseFortran.cmake#L405-L412
Compiling with HDF5 1.12.0 works, and has the following configure output:
-- Performing Test FC_AVAIL_KINDS_RESULT - Success
-- ....NUMBER OF INTEGER KINDS FOUND 4
-- ....REAL KINDS FOUND {4,8}
-- ....INTEGER KINDS FOUND {1,2,4,8}
-- ....MAX DECIMAL PRECISION 15
So it seems the detection logic changed. Best guess would be https://github.com/HDFGroup/hdf5/commit/9b5d9680af8401528bb8c3b6d2b4c3cf30ccec5b
I truly don't understand how REAL KINDS FOUND includes kind=16 for armflang...
tmp$ cat test_fortran_r16.f90
PROGRAM main
USE ISO_C_BINDING
USE, INTRINSIC :: ISO_FORTRAN_ENV, ONLY : stdout=>OUTPUT_UNIT
IMPLICIT NONE
REAL (KIND=16) a
WRITE(stdout, '(I0)') SIZEOF(a)
END
tmp$ armflang test_fortran_r16.f90
F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (test_fortran_r16.f90: 5)
0 inform, 0 warnings, 1 severes, 0 fatal for main
tmp$ echo $?
1
In the spack recipes for hdf5, the AMD compiler team have added this (see: https://github.com/spack/spack/pull/47123)
+ # AOCC does not support _Float16
+ if spec.satisfies("@1.14.4: %aocc"):
+ args.append(self.define("HDF5_ENABLE_NONSTANDARD_FEATURE_FLOAT16", False))
@lrbison - this might help workaround this.
I know I encountered this same issue some months ago, but clean forgot what I was doing at the time so what workaround I used back then.
Actually, that patch is for _float16.. kind=16 is something different - 128b floats.. doh. perhaps a similar fix is available.
in the CMake file for this..
# dnl The output from the above program will be:
# dnl -- LINE 1 -- valid integer kinds (comma separated list)
# dnl -- LINE 2 -- valid real kinds (comma separated list)
# dnl -- LINE 3 -- max decimal precision for reals
# dnl -- LINE 4 -- number of valid integer kinds
# dnl -- LINE 5 -- number of valid real kinds
# dnl -- LINE 6 -- number of valid logical kinds
# dnl -- LINE 7 -- valid logical kinds (comma separated list)
the source code at that point generates this output:
1,2,4,8
4,8,16
15
4
3
4
1,2,4,8
which suggests the source - which compiles without error and is:
PROGRAM FC08_AVAIL_KINDS
USE, INTRINSIC :: ISO_FORTRAN_ENV, ONLY : stdout=>OUTPUT_UNIT, integer_kinds, real_kinds, logical_kinds
IMPLICIT NONE
INTEGER :: ik, jk, k, max_decimal_prec
INTEGER :: num_rkinds, num_ikinds, num_lkinds
! Find integer KINDs
num_ikinds = SIZE(integer_kinds)
DO k = 1, num_ikinds
WRITE(stdout,'(I0)', ADVANCE='NO') integer_kinds(k)
IF(k.NE.num_ikinds)THEN
WRITE(stdout,'(A)',ADVANCE='NO') ','
ELSE
WRITE(stdout,'()')
ENDIF
ENDDO
! Find real KINDs
num_rkinds = SIZE(real_kinds)
max_decimal_prec = 1
prec: DO ik = 2, 36
exp: DO jk = 1, 700
k = SELECTED_REAL_KIND(ik,jk)
IF(k.LT.0) EXIT exp
max_decimal_prec = ik
ENDDO exp
ENDDO prec
DO k = 1, num_rkinds
WRITE(stdout,'(I0)', ADVANCE='NO') real_kinds(k)
IF(k.NE.num_rkinds)THEN
WRITE(stdout,'(A)',ADVANCE='NO') ','
ELSE
WRITE(stdout,'()')
ENDIF
ENDDO
WRITE(stdout,'(I0)') max_decimal_prec
WRITE(stdout,'(I0)') num_ikinds
WRITE(stdout,'(I0)') num_rkinds
! Find logical KINDs
num_lkinds = SIZE(logical_kinds)
WRITE(stdout,'(I0)') num_lkinds
DO k = 1, num_lkinds
WRITE(stdout,'(I0)', ADVANCE='NO') logical_kinds(k)
IF(k.NE.num_lkinds)THEN
WRITE(stdout,'(A)',ADVANCE='NO') ','
ELSE
WRITE(stdout,'()')
ENDIF
ENDDO
END PROGRAM FC08_AVAIL_KINDS
is not getting correct output.
@hyoklee - are you able to reproduce this - do you have access to Arm Compiler for Linux?
What is the output? Depending on how old your compiler is, it could be this issue:
https://github.com/llvm/llvm-project/issues/77282
If the program does not work, then the compiler has a bug.
Thanks @brtnfld - I don't have the F90 chops myself to know what ought to have been the output.
ACfL is built using the old PGI open sourced F90 compiler, as used by NVHPC and AOCC too - but everyone has been working towards flang-new. Arm hasn't migrated to that yet (but has made its flang-new beta available). So, whilst ACfL is new (24.10.1 ) and that's also based on recent LLVM - it's not yet the LLVM flang-new..
Can you suggest a workaround CMake define / flag that would resolve this? I can definitely make a dirty patch into the hdf5 spack recipe that would only apply that with ACfL if there is one.
FWIW :
with ACfL:
program test2
use iso_fortran_env
integer :: i
do i = 1, size(real_kinds)
print *, real_kinds(i)
end do
end program test2
yields: 4, 8, 16.
With the new beta:
ubuntu@ip-10-11-32-85:~$ ./atfl/bin/flang ./test2.f90
ubuntu@ip-10-11-32-85:~$ ./a.out
2
3
4
8
16
FWIW, AMD's AOCC yields 4, 8, 16. I'm not sure if that's still based on the old fortran front end, or if that's correct for that platform.
The output for the second case looks correct to me; it detects the low-precision reals. Could you reiterate what is failing?
The failure is that KIND=16 is not supported by ACfL, but that hdf5 thinks it is.
The failure is that KIND=16 is not supported by ACfL, but that hdf5 thinks it is.
Right, what do you mean by not supported? It is listed as supported in real_kinds from iso_fortran_env?
Does the program
REAL(KIND=16) :: a
END
Not compile?
That's right - it doesn't compile:
[rocky@ip-10-11-42-205 ~]$ module load acfl
Loading acfl/24.10.1
Loading requirement: binutils/14.2.0
[rocky@ip-10-11-42-205 ~]$ armflang test3.f90
F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (test3.f90: 1)
0 inform, 0 warnings, 1 severes, 0 fatal for MAIN
128-bit floats are not supported in hardware on aarch64 - they're software emulated by a GCC library (slowly..), which isn't supported by that fortran compiler.
The older way, then, was to use SELECTED_REAL_KIND Does this print -1, or 16?
k = SELECTED_REAL_KIND(19,2)
print*,k
end
-1.
[rocky@ip-10-11-42-205 ~]$ cat >Test.f90
k = SELECTED_REAL_KIND(19,2)
print*,k
end
[rocky@ip-10-11-42-205 ~]$ module load acfl
Loading acfl/24.10.1
Loading requirement: binutils/14.2.0
[rocky@ip-10-11-42-205 ~]$ armflang Test.f90
[rocky@ip-10-11-42-205 ~]$ ./a.out
-1
Ok, so for your patch, you can add in config/cmake/HDF5UseFortran.cmake the line
set (${HDF_PREFIX}_HAVE_ISO_FORTRAN_ENV 0)
after
# Check if the fortran compiler supports the intrinsic module "ISO_FORTRAN_ENV" (F08)
READ_SOURCE("PROGRAM PROG_FC_ISO_FORTRAN_ENV" "END PROGRAM PROG_FC_ISO_FORTRAN_ENV" SOURCE_CODE)
check_fortran_source_compiles (${SOURCE_CODE} HAVE_ISO_FORTRAN_ENV SRC_EXT f90)
if (${HAVE_ISO_FORTRAN_ENV})
set (${HDF_PREFIX}_HAVE_ISO_FORTRAN_ENV 1)
else ()
set (${HDF_PREFIX}_HAVE_ISO_FORTRAN_ENV 0)
endif ()
So it selects the test that does not use ISO_FORTRAN_ENV
Alas no joy:
==> Using cached archive: /home/rocky/spack/var/spack/cache/_source-cache/archive/ec/ec2e13c52e60f9a01491bb3158cb3778c985697131fc6a342262d32a26e58e44.tar.gz
==> Applied patch /home/rocky/spack/var/spack/repos/builtin/packages/hdf5/acfl.kind16.patch
==> Ran patch() for hdf5
==> hdf5: Executing phase: 'cmake'
==> hdf5: Executing phase: 'build'
==> Error: ProcessError: Command exited with status 2:
'/home/rocky/spack/opt/spack/linux-rocky9-neoverse_v2/arm-24.10.1/gmake-4.4.1-icbfetxuqlpjx4zqfk6kht52fqecbt3e/bin/make' '-j16'
5 errors found in build log:
3134 F90-S-0081-Illegal selector - KIND parameter has unknown value for data type (/tmp/rocky/spack-stage/spack-stage-hdf5-1.14.5-i7uffkyiywiumpumb7zqtehvqi3zxo7c/spack-build-i7uffky/fortran/static/H5_gen.F90:
6049)
where acfl.kind16.patch is:
/home/rocky/spack/var/spack/repos/builtin/packages/hdf5/acfl.kind16.patch
*** a/config/cmake/HDFUseFortran.cmake Wed Mar 12 14:20:33 2025
--- b/config/cmake/HDFUseFortran.cmake Wed Mar 12 14:21:44 2025
*************** endif ()
*** 87,89 ****
--- 87,91 ----
set (CMAKE_EXE_LINKER_FLAGS_DEBUG "/DEBUG" CACHE STRING "flags" FORCE)
endif ()
endif ()
+
+ set (${HDF_PREFIX}_HAVE_ISO_FORTRAN_ENV 0)
.. and looking at the source, it has applied it cleanly.
the text output of cmake is (as expected I think) unchanged:
-- Performing Test FORTRAN_CHAR_ALLOC - Success
-- ....NUMBER OF INTEGER KINDS FOUND 4
-- ....REAL KINDS FOUND {4,8,16}
-- ....INTEGER KINDS FOUND {1,2,4,8}
-- ....MAX DECIMAL PRECISION 15
-- ....LOGICAL KINDS FOUND {1,2,4,8}
-- ....FOUND SIZEOF for REAL KINDs {4,8,8}
-- Found MPI_Fortran: /home/rocky/spack/opt/spack/linux-rocky9-neoverse_v2/arm-24.10.1/openmpi-5.0.6-m7gsvjwqb5rroxnfif76zg5emsf7oskq/lib/libmpi_usempif08.so (found version "3.1")
-- Found MPI: TRUE (found version "3.1") found components: Fortran
The root cause of the issue (as pointed out by @dslarm) is that ACfL is lying about what kinds are supported in iso_fortran_env's real_kinds.
I've made a PR that fixes the issue: https://github.com/HDFGroup/hdf5/pull/5401
Gfortran:
-- Check for working Fortran compiler: /fsx/spack/opt/spack/linux-ubuntu20.04-armv8.4a/gcc-9.4.0/gcc-12.4.0-vfykzopukml5pu7eedkbhcmq22blh25o/bin/gfortran - skipped ...
-- ....NUMBER OF INTEGER KINDS FOUND 5
-- ....REAL KINDS FOUND {4,8,16}
-- ....INTEGER KINDS FOUND {1,2,4,8,16}
-- ....MAX DECIMAL PRECISION 33
-- ....LOGICAL KINDS FOUND {1,2,4,8,16}
-- ....FOUND SIZEOF for REAL KINDs {4,8,16}
ACfL:
-- Check for working Fortran compiler: /fsx/spack/opt/spack/linux-ubuntu20.04-aarch64/gcc-9.4.0/acfl-24.10.1-4uj73pi7jwmdd2w6u2efp674l55hapbv/arm-linux-compiler-24.10.1_Ubuntu-20.04/bin/armflang - skipped
...
-- ....NUMBER OF INTEGER KINDS FOUND 4
-- ....REAL KINDS FOUND {4,8}
-- ....INTEGER KINDS FOUND {1,2,4,8}
-- ....MAX DECIMAL PRECISION 15
-- ....LOGICAL KINDS FOUND {1,2,4,8}
-- ....FOUND SIZEOF for REAL KINDs {4,8}
