ecmwf-ifs / loki

Freely programmable source-to-source translation for Fortran
https://sites.ecmwf.int/docs/loki/
Apache License 2.0
29 stars 12 forks source link

end pragma with curly braces #300

Open ecossevin opened 5 months ago

ecossevin commented 5 months ago

1) ACDC PARALLEL TARGET [...] works but not ACDC PARALLEL,TARGET [...] (see fcode2 vs fcode3) 2)loki doesn't recognize curly braces to define pragma region (see fcode2 vs fcode4)

from loki import * 

fcode1 = """
    SUBROUTINE TEST

    !$loki parallel TARGET=OpenMP/OpenMPSingleColumn/OpenACCSingleColumn,NAME=CPPHINP 
    PRINT *, "loki"
    !$loki end parallel

    END SUBROUTINE"""

routine = Sourcefile.from_source(fcode1)["TEST"]
with pragma_regions_attached(routine):
    for region in FindNodes(PragmaRegion).visit(routine.body):
        print(f"{region=}")
        print("region = ", fgen(region.body))
        assert fgen(region.body) == 'PRINT *, "loki"'

fcode2 = """
    SUBROUTINE TEST

    !$ACDC PARALLEL TARGET=OpenMP/OpenMPSingleColumn/OpenACCSingleColumn,NAME=CPPHINP 
    PRINT *, "ACDC_endparallel"
    !$ACDC END PARALLEL

    END SUBROUTINE"""

routine = Sourcefile.from_source(fcode2)["TEST"]
with pragma_regions_attached(routine):
    for region in FindNodes(PragmaRegion).visit(routine.body):
        print(f"{region=}")
        print("region = ", fgen(region.body))
        assert fgen(region.body) == 'PRINT *, "ACDC_endparallel"'

fcode3 = """
    SUBROUTINE TEST

    !$ACDC PARALLEL,TARGET=OpenMP/OpenMPSingleColumn/OpenACCSingleColumn,NAME=CPPHINP 
    PRINT *, "ACDC_coma"
    !$ACDC END PARALLEL

    END SUBROUTINE"""
routine = Sourcefile.from_source(fcode3)["TEST"]
with pragma_regions_attached(routine):
    for region in FindNodes(PragmaRegion).visit(routine.body):
        print(f"{region=}")
        print("region = ", fgen(region.body))
        assert fgen(region.body) == 'PRINT *, "ACDC_coma"'

fcode4 = """
    SUBROUTINE TEST

    !$ACDC PARALLEL TARGET=OpenMP/OpenMPSingleColumn/OpenACCSingleColumn,NAME=CPPHINP {
    PRINT *, "ACDC_curly_brace"
    !$ACDC }

    END SUBROUTINE"""
routine = Sourcefile.from_source(fcode4)["TEST"]
with pragma_regions_attached(routine):
    for region in FindNodes(PragmaRegion).visit(routine.body):
        print(f"{region=}")
        print("region = ", fgen(region.body))
        assert fgen(region.body) == 'PRINT *, "ACDC_curly_brace"'
ecossevin commented 1 week ago

Hi, Can someone please fix this ? Thanks in advance for your help.

mlange05 commented 1 week ago

Hi @ecossevin , apologies for the delay. I started to look into this, and then realised where the hidden complexities come from, and had to drop it due to time constraints. The basic problem with this notation is that our PragmaRegion matching is based on explicit keyword matching between start and end-pragma, so that we can support nesting different pragma types. For example, we can support

!$loki foo
...
!$loki bar
...
!$loki end bar
!$loki end foo

The particular matching code is here: https://github.com/ecmwf-ifs/loki/blob/main/loki/ir/pragma_utils.py#L467

Unfortunately, re-writing the logic itself to match by order, rather than keyword is tedious. We, of course, accept PRs if you manage to keep this backward compatible yourself. Is there any specific reason why this non-standard pragma notation is required over more standard (eg. similar to OpenACC/OpenMP) notation based on 'end' keywords and spaces over commas?

ecossevin commented 1 week ago

Hi,

Thanks for your answer.

Anyway, I don't need this to work, I have a work around.

There isn't any fondamental reason for that, but we want to keep it like that for the moment.