Hi Tobias! On 2019-10-24T14:47:58+0200, Tobias Burnus <tobias_bur...@mentor.com> wrote: > The clause (new in OpenACC 2.6) makes any device code use the local > memory address for each of the variables specified unless the given > variable is already present on the current device. – Or in words of > OpenACC 2.7 (in Sect. 2.7.9 no_create clause): > > "The no_create clause may appear on structured data and compute > constructs." / "For each var in varlist, if var is in shared memory, no > action is taken; if var is not in shared memory, the no_create clause > behaves as follows:" [digest: if present, update present count, if > pointer attach/detach; if not not present, device-local memory used.]
s%not not%not%
s%device-local%local%
> "The restrictions regarding subarrays in the present clause apply to
> this clause."
> Note: The "no_create" maps to the (new) GOMP_MAP_NO_ALLOC in the middle
> end – and all action in libgomp/target.c but only applies to
> GOMP_MAP_NO_ALLOC; hence, the code should only affect OpenACC.
Not sure if 'GOMP_MAP_NO_ALLOC' is the most descriptive name. ;-)
I understand 'no_create' to mean 'present' in combination with an
'if_present' flag that is available as a clause for some (other) OpenACC
directives, correct?
So, how about naming this 'GOMP_MAP_IF_PRESENT' instead of
'GOMP_MAP_NO_ALLOC'? (Jakub?) (But I don't care too much, so if there's
a good reason to prefer 'GOMP_MAP_NO_ALLOC', then that's fine, too.)
Ah, I just found that Julian (CCed for your information) internally had
proposed 'GOMP_MAP_MAYBE_PRESENT', which seems like another good option
indeed.
For reference:
> --- a/include/gomp-constants.h
> +++ b/include/gomp-constants.h
> @@ -75,6 +75,8 @@ enum gomp_map_kind
> GOMP_MAP_DEVICE_RESIDENT = (GOMP_MAP_FLAG_SPECIAL_1 | 1),
> /* OpenACC link. */
> GOMP_MAP_LINK = (GOMP_MAP_FLAG_SPECIAL_1 | 2),
> + /* Use device data if present, fall back to host address otherwise. */
> + GOMP_MAP_NO_ALLOC = (GOMP_MAP_FLAG_SPECIAL_1 | 3),
> /* Do not map, copy bits for firstprivate instead. */
> GOMP_MAP_FIRSTPRIVATE = (GOMP_MAP_FLAG_SPECIAL | 0),
> /* Similarly, but store the value in the pointer rather than
> OK for the trunk?
To synchronize our efforts, I'm attaching an incremental WIP patch. Will
you please have a look at that, merging it in, while I continue to
review?
> PS: This patch is a re-diffed version of the OG9/OG8 version; as some
> other features are not yet on trunk, it misses a test case for
> "no_create(s.y…)" (i.e. the struct component-ref;
> libgomp/testsuite/libgomp.oacc-c-c++-common/nocreate-{3,4}.c); trunk
> also lacks 'acc serial' and, hence, the attach patch lacks the
> OACC_SERIAL_CLAUSE_MASK updates – and gfc_match_omp_map_clause needs
> later to be updated for the allow_derived and allow_common arguments.
> Furthermore, some 'do_detach = false' are missing in libgomp/target.c as
> they do not yet exist on trunk, either.
>
> The openacc-gcc-9 /…-8 branch patch is commit
> 8e74c2ec2b90819c995444370e742864a685209f of Dec 20, 2018. It has been
> posted as https://gcc.gnu.org/ml/gcc-patches/2018-12/msg01418.html
Thanks for providing these references, that's useful.
> libgomp/
> * testsuite/libgomp.oacc-c-c++-common/nocreate-1.c: New test.
> * testsuite/libgomp.oacc-c-c++-common/nocreate-2.c: New test.
> * testsuite/libgomp.oacc-fortran/nocreate-1.f90: New test.
> * testsuite/libgomp.oacc-fortran/nocreate-2.f90: New test.
Please rename these files to 'no_create*', as that's what the clause is
called.
..., and then:
> --- a/libgomp/target.c
> +++ b/libgomp/target.c
> @@ -667,6 +667,12 @@ gomp_map_vars_internal (struct gomp_device_descr
> *devicep,
> has_firstprivate = true;
> continue;
> }
> + else if ((kind & typemask) == GOMP_MAP_NO_ALLOC)
> + {
> + tgt->list[i].key = NULL;
> + tgt->list[i].offset = 0;
> + continue;
> + }
> cur_node.host_start = (uintptr_t) hostaddrs[i];
> if (!GOMP_MAP_POINTER_P (kind & typemask))
> cur_node.host_end = cur_node.host_start + sizes[i];
> @@ -892,6 +898,49 @@ gomp_map_vars_internal (struct gomp_device_descr
> *devicep,
> cur_node.tgt_offset = n->tgt->tgt_start + n->tgt_offset
> + cur_node.host_start - n->host_start;
> continue;
> + case GOMP_MAP_NO_ALLOC:
> + {
> + cur_node.host_start = (uintptr_t) hostaddrs[i];
> + cur_node.host_end = cur_node.host_start + sizes[i];
> + splay_tree_key n = splay_tree_lookup (mem_map, &cur_node);
> + if (n != NULL)
> + {
> + tgt->list[i].key = n;
> + tgt->list[i].offset = cur_node.host_start - n->host_start;
> + tgt->list[i].length = n->host_end - n->host_start;
> + tgt->list[i].copy_from = false;
> + tgt->list[i].always_copy_from = false;
> + n->refcount++;
> + }
> + else
> + {
> + tgt->list[i].key = NULL;
> + tgt->list[i].offset = OFFSET_INLINED;
> + tgt->list[i].length = sizes[i];
> + tgt->list[i].copy_from = false;
> + tgt->list[i].always_copy_from = false;
> + if (i + 1 < mapnum)
> + {
> + int kind2 = get_kind (short_mapkind, kinds, i + 1);
> + switch (kind2 & typemask)
> + {
> + case GOMP_MAP_POINTER:
> + /* The data is not present but we have an attach
> + or pointer clause next. Skip over it. */
> + i++;
> + tgt->list[i].key = NULL;
> + tgt->list[i].offset = OFFSET_INLINED;
> + tgt->list[i].length = sizes[i];
> + tgt->list[i].copy_from = false;
> + tgt->list[i].always_copy_from = false;
> + break;
> + default:
> + break;
> + }
> + }
> + }
> + continue;
> + }
> default:
> break;
> }
This I don't grok yet; see the "TODO" comments in the attached
incremental WIP patch.
Grüße
Thomas
From 22ceeb89f787a6279a37d35965f82a4f5b3e3b72 Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <tho...@codesourcery.com>
Date: Wed, 6 Nov 2019 00:42:06 +0100
Subject: [PATCH] [WIP] into Add OpenACC 2.6 `no_create' clause support
---
gcc/fortran/openmp.c | 4 ++--
.../gfortran.dg/goacc/common-block-1.f90 | 3 +++
.../gfortran.dg/goacc/common-block-2.f90 | 3 +++
.../gfortran.dg/goacc/data-clauses.f95 | 23 ++++++++++++++++++-
gcc/testsuite/gfortran.dg/goacc/data-tree.f95 | 3 ++-
.../gfortran.dg/goacc/kernels-tree.f95 | 3 ++-
.../gfortran.dg/goacc/parallel-tree.f95 | 3 ++-
libgomp/target.c | 8 +++++++
.../libgomp.oacc-fortran/common-block-2.f90 | 4 +++-
.../libgomp.oacc-fortran/nocreate-1.f90 | 10 ++++++--
.../libgomp.oacc-fortran/nocreate-2.f90 | 16 ++++++++++---
11 files changed, 68 insertions(+), 12 deletions(-)
diff --git a/gcc/fortran/openmp.c b/gcc/fortran/openmp.c
index 47c5cf5d422..822af5dbe7c 100644
--- a/gcc/fortran/openmp.c
+++ b/gcc/fortran/openmp.c
@@ -1449,7 +1449,7 @@ gfc_match_omp_clauses (gfc_omp_clauses **cp, const omp_mask mask,
if ((mask & OMP_CLAUSE_NO_CREATE)
&& gfc_match ("no_create ( ") == MATCH_YES
&& gfc_match_omp_map_clause (&c->lists[OMP_LIST_MAP],
- OMP_MAP_NO_ALLOC))
+ OMP_MAP_NO_ALLOC, true))
continue;
if ((mask & OMP_CLAUSE_NOGROUP)
&& !c->nogroup
@@ -1969,7 +1969,7 @@ gfc_match_omp_clauses (gfc_omp_clauses **cp, const omp_mask mask,
| OMP_CLAUSE_NUM_WORKERS | OMP_CLAUSE_VECTOR_LENGTH | OMP_CLAUSE_DEVICEPTR \
| OMP_CLAUSE_COPY | OMP_CLAUSE_COPYIN | OMP_CLAUSE_COPYOUT \
| OMP_CLAUSE_CREATE | OMP_CLAUSE_NO_CREATE | OMP_CLAUSE_PRESENT \
- | OMP_CLAUSE_DEFAULT | OMP_CLAUSE_WAIT)
+ | OMP_CLAUSE_DEFAULT | OMP_CLAUSE_WAIT)
#define OACC_DATA_CLAUSES \
(omp_mask (OMP_CLAUSE_IF) | OMP_CLAUSE_DEVICEPTR | OMP_CLAUSE_COPY \
| OMP_CLAUSE_COPYIN | OMP_CLAUSE_COPYOUT | OMP_CLAUSE_CREATE \
diff --git a/gcc/testsuite/gfortran.dg/goacc/common-block-1.f90 b/gcc/testsuite/gfortran.dg/goacc/common-block-1.f90
index ea437526b46..5c162a5b884 100644
--- a/gcc/testsuite/gfortran.dg/goacc/common-block-1.f90
+++ b/gcc/testsuite/gfortran.dg/goacc/common-block-1.f90
@@ -51,6 +51,9 @@ program test
!$acc data pcopyout(/blockA/, /blockB/, e, v)
!$acc end data
+ !$acc data no_create(/blockA/, /blockB/, e, v)
+ !$acc end data
+
!$acc parallel private(/blockA/, /blockB/, e, v)
!$acc end parallel
diff --git a/gcc/testsuite/gfortran.dg/goacc/common-block-2.f90 b/gcc/testsuite/gfortran.dg/goacc/common-block-2.f90
index 1ba945019f9..33c0d3f5fb4 100644
--- a/gcc/testsuite/gfortran.dg/goacc/common-block-2.f90
+++ b/gcc/testsuite/gfortran.dg/goacc/common-block-2.f90
@@ -39,6 +39,9 @@ program test
!$acc data pcopyout(/blockA/, /blockB/, e, v, a) ! { dg-error "Symbol .a. present on multiple clauses" }
!$acc end data
+ !$acc data no_create(/blockA/, /blockB/, e, v, a) ! { dg-error "Symbol .a. present on multiple clauses" }
+ !$acc end data
+
!$acc parallel private(/blockA/, /blockB/, e, v, a) ! { dg-error "Symbol .a. present on multiple clauses" }
!$acc end parallel
diff --git a/gcc/testsuite/gfortran.dg/goacc/data-clauses.f95 b/gcc/testsuite/gfortran.dg/goacc/data-clauses.f95
index b94214e8b63..c1b3e1dec38 100644
--- a/gcc/testsuite/gfortran.dg/goacc/data-clauses.f95
+++ b/gcc/testsuite/gfortran.dg/goacc/data-clauses.f95
@@ -111,6 +111,27 @@ contains
!$acc end data
+ !$acc parallel no_create (tip) ! { dg-error "POINTER" }
+ !$acc end parallel
+ !$acc parallel no_create (tia) ! { dg-error "ALLOCATABLE" }
+ !$acc end parallel
+ !$acc parallel deviceptr (i) no_create (i) ! { dg-error "multiple clauses" }
+ !$acc end parallel
+ !$acc parallel copy (i) no_create (i) ! { dg-error "multiple clauses" }
+ !$acc end parallel
+ !$acc parallel copyin (i) no_create (i) ! { dg-error "multiple clauses" }
+ !$acc end parallel
+ !$acc parallel copyout (i) no_create (i) ! { dg-error "multiple clauses" }
+ !$acc end parallel
+
+ !$acc parallel no_create (i, c, r, ia, ca, ra, asa, rp, ti, vi, aa)
+ !$acc end parallel
+ !$acc kernels no_create (i, c, r, ia, ca, ra, asa, rp, ti, vi, aa)
+ !$acc end kernels
+ !$acc data no_create (i, c, r, ia, ca, ra, asa, rp, ti, vi, aa)
+ !$acc end data
+
+
!$acc parallel present (tip) ! { dg-error "POINTER" }
!$acc end parallel
!$acc parallel present (tia) ! { dg-error "ALLOCATABLE" }
@@ -256,4 +277,4 @@ contains
!$acc end data
end subroutine foo
-end module test
\ No newline at end of file
+end module test
diff --git a/gcc/testsuite/gfortran.dg/goacc/data-tree.f95 b/gcc/testsuite/gfortran.dg/goacc/data-tree.f95
index f16d62cce69..454417d6a05 100644
--- a/gcc/testsuite/gfortran.dg/goacc/data-tree.f95
+++ b/gcc/testsuite/gfortran.dg/goacc/data-tree.f95
@@ -7,6 +7,7 @@ program test
logical :: l = .true.
!$acc data if(l) copy(i), copyin(j), copyout(k), create(m) &
+ !$acc no_create(n) &
!$acc present(o), pcopy(p), pcopyin(r), pcopyout(s), pcreate(t) &
!$acc deviceptr(u)
!$acc end data
@@ -19,7 +20,7 @@ end program test
! { dg-final { scan-tree-dump-times "map\\(to:j\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(from:k\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(alloc:m\\)" 1 "original" } }
-
+! { dg-final { scan-tree-dump-times "map\\(no_alloc:n\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(force_present:o\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(tofrom:p\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(to:r\\)" 1 "original" } }
diff --git a/gcc/testsuite/gfortran.dg/goacc/kernels-tree.f95 b/gcc/testsuite/gfortran.dg/goacc/kernels-tree.f95
index a70f1e737bd..5583ffb4d04 100644
--- a/gcc/testsuite/gfortran.dg/goacc/kernels-tree.f95
+++ b/gcc/testsuite/gfortran.dg/goacc/kernels-tree.f95
@@ -8,6 +8,7 @@ program test
!$acc kernels if(l) async num_gangs(i) num_workers(i) vector_length(i) &
!$acc copy(i), copyin(j), copyout(k), create(m) &
+ !$acc no_create(n) &
!$acc present(o), pcopy(p), pcopyin(r), pcopyout(s), pcreate(t) &
!$acc deviceptr(u)
!$acc end kernels
@@ -25,7 +26,7 @@ end program test
! { dg-final { scan-tree-dump-times "map\\(to:j\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(from:k\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(alloc:m\\)" 1 "original" } }
-
+! { dg-final { scan-tree-dump-times "map\\(no_alloc:n\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(force_present:o\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(tofrom:p\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(to:r\\)" 1 "original" } }
diff --git a/gcc/testsuite/gfortran.dg/goacc/parallel-tree.f95 b/gcc/testsuite/gfortran.dg/goacc/parallel-tree.f95
index 2697bb79e7f..e33653bdd78 100644
--- a/gcc/testsuite/gfortran.dg/goacc/parallel-tree.f95
+++ b/gcc/testsuite/gfortran.dg/goacc/parallel-tree.f95
@@ -9,6 +9,7 @@ program test
!$acc parallel if(l) async num_gangs(i) num_workers(i) vector_length(i) &
!$acc reduction(max:q), copy(i), copyin(j), copyout(k), create(m) &
+ !$acc no_create(n) &
!$acc present(o), pcopy(p), pcopyin(r), pcopyout(s), pcreate(t) &
!$acc deviceptr(u), private(v), firstprivate(w)
!$acc end parallel
@@ -28,7 +29,7 @@ end program test
! { dg-final { scan-tree-dump-times "map\\(to:j\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(from:k\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(alloc:m\\)" 1 "original" } }
-
+! { dg-final { scan-tree-dump-times "map\\(no_alloc:n\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(force_present:o\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(tofrom:p\\)" 1 "original" } }
! { dg-final { scan-tree-dump-times "map\\(to:r\\)" 1 "original" } }
diff --git a/libgomp/target.c b/libgomp/target.c
index 632e7020538..0338648946d 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -669,6 +669,7 @@ gomp_map_vars_internal (struct gomp_device_descr *devicep,
}
else if ((kind & typemask) == GOMP_MAP_NO_ALLOC)
{
+ //TODO TS is confused. Handling this here, will inhibit 'gomp_map_vars_existing' being used a bit further below.
tgt->list[i].key = NULL;
tgt->list[i].offset = 0;
continue;
@@ -905,6 +906,7 @@ gomp_map_vars_internal (struct gomp_device_descr *devicep,
splay_tree_key n = splay_tree_lookup (mem_map, &cur_node);
if (n != NULL)
{
+ //TODO TS is confused. Due to the way the handling of 'GOMP_MAP_NO_ALLOC' is done in the first loop, we're here re-doing 'gomp_map_vars_existing'?
tgt->list[i].key = n;
tgt->list[i].offset = cur_node.host_start - n->host_start;
tgt->list[i].length = n->host_end - n->host_start;
@@ -914,6 +916,7 @@ gomp_map_vars_internal (struct gomp_device_descr *devicep,
}
else
{
+ //TODO This is basically 'GOMP_MAP_FIRSTPRIVATE_INT' handling?
tgt->list[i].key = NULL;
tgt->list[i].offset = OFFSET_INLINED;
tgt->list[i].length = sizes[i];
@@ -925,6 +928,11 @@ gomp_map_vars_internal (struct gomp_device_descr *devicep,
switch (kind2 & typemask)
{
case GOMP_MAP_POINTER:
+ //TODO abort();
+ //TODO This code path is exercised by 'libgomp.oacc-fortran/nocreate-2.f90'.
+ //TODO TS does not yet understand why this is needed.
+ //TODO Is this somehow similar to 'GOMP_MAP_TO_PSET' handling?
+
/* The data is not present but we have an attach
or pointer clause next. Skip over it. */
i++;
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/common-block-2.f90 b/libgomp/testsuite/libgomp.oacc-fortran/common-block-2.f90
index 018b37d00bb..ad04ca997c2 100644
--- a/libgomp/testsuite/libgomp.oacc-fortran/common-block-2.f90
+++ b/libgomp/testsuite/libgomp.oacc-fortran/common-block-2.f90
@@ -76,7 +76,9 @@ program main
!$acc enter data create(b)
- !$acc parallel loop pcopy(b)
+ !$acc parallel loop &
+ !$acc no_create(b) ! ... here means 'present(b)'.
+ !TODO But we get: "libgomp: cuStreamSynchronize error: an illegal memory access was encountered".
do i = 1, n
b(i) = i
end do
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/nocreate-1.f90 b/libgomp/testsuite/libgomp.oacc-fortran/nocreate-1.f90
index f048355d7df..ca9611b777c 100644
--- a/libgomp/testsuite/libgomp.oacc-fortran/nocreate-1.f90
+++ b/libgomp/testsuite/libgomp.oacc-fortran/nocreate-1.f90
@@ -1,20 +1,26 @@
-! { dg-skip-if "" { *-*-* } { "*" } { "-DACC_MEM_SHARED=0" } }
+! { dg-do run }
! Test no_create clause with data construct when data is present/not present.
program nocreate
use openacc
implicit none
+ logical :: shared_memory
integer, parameter :: n = 512
integer :: myarr(n)
integer i
+ shared_memory = .false.
+ !$acc kernels copyin (shared_memory)
+ shared_memory = .true.
+ !$acc end kernels
+
do i = 1, n
myarr(i) = 0
end do
!$acc data no_create (myarr)
- if (acc_is_present (myarr)) stop 1
+ if (acc_is_present (myarr) .neqv. shared_memory) stop 1
!$acc end data
!$acc enter data copyin (myarr)
diff --git a/libgomp/testsuite/libgomp.oacc-fortran/nocreate-2.f90 b/libgomp/testsuite/libgomp.oacc-fortran/nocreate-2.f90
index 34444ecf5b0..16227b8ae22 100644
--- a/libgomp/testsuite/libgomp.oacc-fortran/nocreate-2.f90
+++ b/libgomp/testsuite/libgomp.oacc-fortran/nocreate-2.f90
@@ -1,14 +1,20 @@
-! { dg-skip-if "" { *-*-* } { "*" } { "-DACC_MEM_SHARED=0" } }
+! { dg-do run }
! Test no_create clause with data/parallel constructs.
program nocreate
use openacc
implicit none
+ logical :: shared_memory
integer, parameter :: n = 512
integer :: myarr(n)
integer i
+ shared_memory = .false.
+ !$acc kernels copyin (shared_memory)
+ shared_memory = .true.
+ !$acc end kernels
+
do i = 1, n
myarr(i) = 0
end do
@@ -16,7 +22,11 @@ program nocreate
call do_on_target(myarr, n)
do i = 1, n
- if (myarr(i) .ne. i) stop 1
+ if (shared_memory) then
+ if (myarr(i) .ne. i * 2) stop 1
+ else
+ if (myarr(i) .ne. i) stop 2
+ end if
end do
do i = 1, n
@@ -28,7 +38,7 @@ program nocreate
!$acc exit data copyout(myarr)
do i = 1, n
- if (myarr(i) .ne. i * 2) stop 2
+ if (myarr(i) .ne. i * 2) stop 3
end do
end program nocreate
--
2.17.1
signature.asc
Description: PGP signature
