* [Patch] libgomp: Fix reverse offload issues
@ 2023-02-02 14:13 Tobias Burnus
2023-02-03 10:38 ` Tobias Burnus
0 siblings, 1 reply; 3+ messages in thread
From: Tobias Burnus @ 2023-02-02 14:13 UTC (permalink / raw)
To: gcc-patches, Jakub Jelinek
[-- Attachment #1: Type: text/plain, Size: 526 bytes --]
Found when testing AMD GCN offloading, the second issue came up with
libgomp.fortran/reverse-offload-5.f90. (But oddly not with nvptx.)
While the first one (new test: libgomp.fortran/reverse-offload-6.f90)
came up when debugging the issue.
Tobias
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
[-- Attachment #2: fix-rev-off.diff --]
[-- Type: text/x-patch, Size: 3310 bytes --]
libgomp: Fix reverse offload issues
If there is nothing to map, skip the mapping and attempting to copy
addrs, sizes and kinds which may have issues for size = 0.
Additionally, it could happen that a non-allocated address was deallocated,
e.g. a pointer set - such that there was a double free for the actual data
or in multiple other ways.
libgomp/
* target.c (gomp_target_rev): Handle mapnum == 0 and avoid
freeing not allocated memory.
* testsuite/libgomp.fortran/reverse-offload-6.f90: New test.
libgomp/target.c | 8 +++---
.../libgomp.fortran/reverse-offload-6.f90 | 32 ++++++++++++++++++++++
2 files changed, 36 insertions(+), 4 deletions(-)
diff --git a/libgomp/target.c b/libgomp/target.c
index b16ee761a95..c1682caea13 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -3324,7 +3324,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
gomp_fatal ("Cannot find reverse-offload function");
void (*host_fn)() = (void (*)()) n->k->host_start;
- if (devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
+ if ((devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) || mapnum == 0)
{
devaddrs = (uint64_t *) (uintptr_t) devaddrs_ptr;
sizes = (uint64_t *) (uintptr_t) sizes_ptr;
@@ -3402,7 +3402,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
}
}
- if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
+ if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
{
size_t j, struct_cpy = 0;
splay_tree_key n2;
@@ -3638,7 +3638,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
host_fn (devaddrs);
- if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
+ if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
{
uint64_t struct_cpy = 0;
bool clean_struct = false;
@@ -3680,7 +3680,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
clean_struct = true;
struct_cpy = sizes[i];
}
- else if (cdata[i].aligned)
+ else if (!cdata[i].present && cdata[i].aligned)
gomp_aligned_free ((void *) (uintptr_t) devaddrs[i]);
else if (!cdata[i].present)
free ((void *) (uintptr_t) devaddrs[i]);
diff --git a/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90 b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
new file mode 100644
index 00000000000..04866edbba7
--- /dev/null
+++ b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
@@ -0,0 +1,32 @@
+!
+! Ensure that a mapping with no argument works
+!
+
+module m
+ implicit none (type, external)
+ integer :: x = 32
+ integer :: dev_num2 = -1
+contains
+subroutine foo()
+ use omp_lib, only: omp_get_device_num
+ x = x + 10
+ dev_num2 = omp_get_device_num()
+end
+end module m
+
+use m
+use omp_lib
+!$omp requires reverse_offload
+implicit none (type, external)
+integer :: dev_num = -1
+!$omp target map(from:dev_num)
+ dev_num = omp_get_device_num()
+ ! This calls GOMP_target_ext with number of maps = 0
+ !$omp target device(ancestor:1)
+ call foo
+ !$omp end target
+!$omp end target
+
+if (omp_get_num_devices() > 0 .and. dev_num2 == dev_num) stop 1
+if (x /= 42) stop 2
+end
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [Patch] libgomp: Fix reverse offload issues
2023-02-02 14:13 [Patch] libgomp: Fix reverse offload issues Tobias Burnus
@ 2023-02-03 10:38 ` Tobias Burnus
2023-02-07 22:47 ` Fix 'libgomp.fortran/reverse-offload-6.f90' nvptx offloading compilation (was: [Patch] libgomp: Fix reverse offload issues) Thomas Schwinge
0 siblings, 1 reply; 3+ messages in thread
From: Tobias Burnus @ 2023-02-03 10:38 UTC (permalink / raw)
To: gcc-patches, Jakub Jelinek
[-- Attachment #1: Type: text/plain, Size: 1352 bytes --]
Now committed as obvious as
r13-5680-g0b1ce70a813b98ef2893779d14ad6c90c5d06a71.
I improved the wording in the commit comment a bit, compared to previous
attachment and I have verified that those features work with AMDGCN* and
without offloading.
Tobias
(* it seems as if there is still another issue with mapping, this time
for array-descriptor variables that do not exist on the host, i.e. that
have to be fully mapped to the host. I will look into this today or Monday.)
On 02.02.23 15:13, Tobias Burnus wrote:
> Found when testing AMD GCN offloading, the second issue came up with
> libgomp.fortran/reverse-offload-5.f90. (But oddly not with nvptx.)
>
> While the first one (new test: libgomp.fortran/reverse-offload-6.f90)
> came up when debugging the issue.
>
> Tobias
> -----------------
> Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße
> 201, 80634 München; Gesellschaft mit beschränkter Haftung;
> Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft:
> München; Registergericht München, HRB 106955
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
[-- Attachment #2: committed.diff --]
[-- Type: text/x-patch, Size: 3461 bytes --]
commit 0b1ce70a813b98ef2893779d14ad6c90c5d06a71
Author: Tobias Burnus <tobias@codesourcery.com>
Date: Fri Feb 3 11:31:53 2023 +0100
libgomp: Fix reverse offload issues
If there is nothing to map, skip the mapping and avoid attempting to
copy 0 bytes from addrs, sizes and kinds.
Additionally, it could happen that a non-allocated address was deallocated,
such as a pointer set, leading to a free for the actual data.
libgomp/
* target.c (gomp_target_rev): Handle mapnum == 0 and avoid
freeing not allocated memory.
* testsuite/libgomp.fortran/reverse-offload-6.f90: New test.
---
libgomp/target.c | 8 +++---
.../libgomp.fortran/reverse-offload-6.f90 | 32 ++++++++++++++++++++++
2 files changed, 36 insertions(+), 4 deletions(-)
diff --git a/libgomp/target.c b/libgomp/target.c
index b16ee761a95..c1682caea13 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -3324,7 +3324,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
gomp_fatal ("Cannot find reverse-offload function");
void (*host_fn)() = (void (*)()) n->k->host_start;
- if (devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
+ if ((devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) || mapnum == 0)
{
devaddrs = (uint64_t *) (uintptr_t) devaddrs_ptr;
sizes = (uint64_t *) (uintptr_t) sizes_ptr;
@@ -3402,7 +3402,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
}
}
- if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
+ if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
{
size_t j, struct_cpy = 0;
splay_tree_key n2;
@@ -3638,7 +3638,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
host_fn (devaddrs);
- if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
+ if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
{
uint64_t struct_cpy = 0;
bool clean_struct = false;
@@ -3680,7 +3680,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
clean_struct = true;
struct_cpy = sizes[i];
}
- else if (cdata[i].aligned)
+ else if (!cdata[i].present && cdata[i].aligned)
gomp_aligned_free ((void *) (uintptr_t) devaddrs[i]);
else if (!cdata[i].present)
free ((void *) (uintptr_t) devaddrs[i]);
diff --git a/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90 b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
new file mode 100644
index 00000000000..04866edbba7
--- /dev/null
+++ b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
@@ -0,0 +1,32 @@
+!
+! Ensure that a mapping with no argument works
+!
+
+module m
+ implicit none (type, external)
+ integer :: x = 32
+ integer :: dev_num2 = -1
+contains
+subroutine foo()
+ use omp_lib, only: omp_get_device_num
+ x = x + 10
+ dev_num2 = omp_get_device_num()
+end
+end module m
+
+use m
+use omp_lib
+!$omp requires reverse_offload
+implicit none (type, external)
+integer :: dev_num = -1
+!$omp target map(from:dev_num)
+ dev_num = omp_get_device_num()
+ ! This calls GOMP_target_ext with number of maps = 0
+ !$omp target device(ancestor:1)
+ call foo
+ !$omp end target
+!$omp end target
+
+if (omp_get_num_devices() > 0 .and. dev_num2 == dev_num) stop 1
+if (x /= 42) stop 2
+end
^ permalink raw reply [flat|nested] 3+ messages in thread
* Fix 'libgomp.fortran/reverse-offload-6.f90' nvptx offloading compilation (was: [Patch] libgomp: Fix reverse offload issues)
2023-02-03 10:38 ` Tobias Burnus
@ 2023-02-07 22:47 ` Thomas Schwinge
0 siblings, 0 replies; 3+ messages in thread
From: Thomas Schwinge @ 2023-02-07 22:47 UTC (permalink / raw)
To: Tobias Burnus, gcc-patches; +Cc: Jakub Jelinek
[-- Attachment #1: Type: text/plain, Size: 5461 bytes --]
Hi Tobias!
On 2023-02-03T11:38:29+0100, Tobias Burnus <tobias@codesourcery.com> wrote:
> Now committed as obvious as
> r13-5680-g0b1ce70a813b98ef2893779d14ad6c90c5d06a71.
I see:
FAIL: libgomp.fortran/reverse-offload-6.f90 -O (test for excess errors)
... due to:
nvptx mkoffload: warning: 'omp requires reverse_offload' requires at least 'sm_35' for '-foffload-options=nvptx-none=-march=' - disabling offload-code generation for this device type
I've pushed to master branch
commit 7ab75a6e6dfd81c566f848ef0d8d06666ee15b81
"Fix 'libgomp.fortran/reverse-offload-6.f90' nvptx offloading compilation",
see attached.
Grüße
Thomas
> I improved the wording in the commit comment a bit, compared to previous
> attachment and I have verified that those features work with AMDGCN* and
> without offloading.
>
> Tobias
>
> (* it seems as if there is still another issue with mapping, this time
> for array-descriptor variables that do not exist on the host, i.e. that
> have to be fully mapped to the host. I will look into this today or Monday.)
>
> On 02.02.23 15:13, Tobias Burnus wrote:
>> Found when testing AMD GCN offloading, the second issue came up with
>> libgomp.fortran/reverse-offload-5.f90. (But oddly not with nvptx.)
>>
>> While the first one (new test: libgomp.fortran/reverse-offload-6.f90)
>> came up when debugging the issue.
>>
>> Tobias
> commit 0b1ce70a813b98ef2893779d14ad6c90c5d06a71
> Author: Tobias Burnus <tobias@codesourcery.com>
> Date: Fri Feb 3 11:31:53 2023 +0100
>
> libgomp: Fix reverse offload issues
>
> If there is nothing to map, skip the mapping and avoid attempting to
> copy 0 bytes from addrs, sizes and kinds.
>
> Additionally, it could happen that a non-allocated address was deallocated,
> such as a pointer set, leading to a free for the actual data.
>
> libgomp/
> * target.c (gomp_target_rev): Handle mapnum == 0 and avoid
> freeing not allocated memory.
> * testsuite/libgomp.fortran/reverse-offload-6.f90: New test.
> ---
> libgomp/target.c | 8 +++---
> .../libgomp.fortran/reverse-offload-6.f90 | 32 ++++++++++++++++++++++
> 2 files changed, 36 insertions(+), 4 deletions(-)
>
> diff --git a/libgomp/target.c b/libgomp/target.c
> index b16ee761a95..c1682caea13 100644
> --- a/libgomp/target.c
> +++ b/libgomp/target.c
> @@ -3324,7 +3324,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
> gomp_fatal ("Cannot find reverse-offload function");
> void (*host_fn)() = (void (*)()) n->k->host_start;
>
> - if (devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
> + if ((devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) || mapnum == 0)
> {
> devaddrs = (uint64_t *) (uintptr_t) devaddrs_ptr;
> sizes = (uint64_t *) (uintptr_t) sizes_ptr;
> @@ -3402,7 +3402,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
> }
> }
>
> - if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
> + if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
> {
> size_t j, struct_cpy = 0;
> splay_tree_key n2;
> @@ -3638,7 +3638,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
>
> host_fn (devaddrs);
>
> - if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM))
> + if (!(devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM) && mapnum > 0)
> {
> uint64_t struct_cpy = 0;
> bool clean_struct = false;
> @@ -3680,7 +3680,7 @@ gomp_target_rev (uint64_t fn_ptr, uint64_t mapnum, uint64_t devaddrs_ptr,
> clean_struct = true;
> struct_cpy = sizes[i];
> }
> - else if (cdata[i].aligned)
> + else if (!cdata[i].present && cdata[i].aligned)
> gomp_aligned_free ((void *) (uintptr_t) devaddrs[i]);
> else if (!cdata[i].present)
> free ((void *) (uintptr_t) devaddrs[i]);
> diff --git a/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90 b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
> new file mode 100644
> index 00000000000..04866edbba7
> --- /dev/null
> +++ b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
> @@ -0,0 +1,32 @@
> +!
> +! Ensure that a mapping with no argument works
> +!
> +
> +module m
> + implicit none (type, external)
> + integer :: x = 32
> + integer :: dev_num2 = -1
> +contains
> +subroutine foo()
> + use omp_lib, only: omp_get_device_num
> + x = x + 10
> + dev_num2 = omp_get_device_num()
> +end
> +end module m
> +
> +use m
> +use omp_lib
> +!$omp requires reverse_offload
> +implicit none (type, external)
> +integer :: dev_num = -1
> +!$omp target map(from:dev_num)
> + dev_num = omp_get_device_num()
> + ! This calls GOMP_target_ext with number of maps = 0
> + !$omp target device(ancestor:1)
> + call foo
> + !$omp end target
> +!$omp end target
> +
> +if (omp_get_num_devices() > 0 .and. dev_num2 == dev_num) stop 1
> +if (x /= 42) stop 2
> +end
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0001-Fix-libgomp.fortran-reverse-offload-6.f90-nvptx-offl.patch --]
[-- Type: text/x-diff, Size: 1094 bytes --]
From 7ab75a6e6dfd81c566f848ef0d8d06666ee15b81 Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <thomas@codesourcery.com>
Date: Tue, 7 Feb 2023 21:23:47 +0100
Subject: [PATCH] Fix 'libgomp.fortran/reverse-offload-6.f90' nvptx offloading
compilation
Fix-up for recent commit 0b1ce70a813b98ef2893779d14ad6c90c5d06a71
"libgomp: Fix reverse offload issues".
libgomp/
* testsuite/libgomp.fortran/reverse-offload-6.f90: Fix nvptx
offloading compilation.
---
libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90 | 2 ++
1 file changed, 2 insertions(+)
diff --git a/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90 b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
index 04866edbba7d..ddb7008cb1ad 100644
--- a/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
+++ b/libgomp/testsuite/libgomp.fortran/reverse-offload-6.f90
@@ -2,6 +2,8 @@
! Ensure that a mapping with no argument works
!
+! { dg-additional-options -foffload-options=nvptx-none=-misa=sm_35 { target offload_target_nvptx } }
+
module m
implicit none (type, external)
integer :: x = 32
--
2.39.0
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2023-02-07 22:47 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-02-02 14:13 [Patch] libgomp: Fix reverse offload issues Tobias Burnus
2023-02-03 10:38 ` Tobias Burnus
2023-02-07 22:47 ` Fix 'libgomp.fortran/reverse-offload-6.f90' nvptx offloading compilation (was: [Patch] libgomp: Fix reverse offload issues) Thomas Schwinge
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).