On Thu, 14 Nov 2019 12:08:45 +0100
Thomas Schwinge wrote:
> Hi!
>
> In context of reviewing Julian's "OpenACC reference count overhaul",
> I'm generally reviewing (also known as: trying to understand) the
> libgomp OpenMP 'target' "refcount"ing , and I noticed something
> strange (?):
>
> On
On Thu, Nov 14, 2019 at 12:08:45PM +0100, Thomas Schwinge wrote:
> > @@ -794,7 +807,7 @@ gomp_offload_image_to_device (struct gomp_device_descr
> > *devicep,
> >/* Insert host-target address mapping into splay tree. */
> >struct target_mem_desc *tgt = gomp_malloc (sizeof (*tgt));
> >
Hi!
In context of reviewing Julian's "OpenACC reference count overhaul", I'm
generally reviewing (also known as: trying to understand) the libgomp
OpenMP 'target' "refcount"ing , and I noticed something strange (?):
On 2015-07-30T22:44:33+0300, Ilya Verbin wrote:
> make check-target-libgomp
On Thu, Jul 30, 2015 at 05:40:15PM +0300, Ilya Verbin wrote:
Sure, but it's not possible to fully test data mapping without non-shared
address space. I've created new check_effective_target, ok for
gomp-4_1-branch?
* testsuite/lib/libgomp.exp
On Thu, Jul 30, 2015 at 10:12:59 +0200, Jakub Jelinek wrote:
This test will fail on HSA, you don't assume just that it doesn't
fallback to host, but also non-shared address space.
I think it would be better to start with some check for non-shared address
space, like:
/* This test relies on
On Thu, Jul 30, 2015 at 10:12:59 +0200, Jakub Jelinek wrote:
On Wed, Jul 29, 2015 at 10:06:52PM +0300, Ilya Verbin wrote:
@@ -1241,6 +1245,62 @@ GOMP_target_update (int device, const void *unused,
size_t mapnum,
gomp_update (devicep, mapnum, hostaddrs, sizes, kinds, false);
}
On Thu, Jul 30, 2015 at 10:44:33PM +0300, Ilya Verbin wrote:
libgomp/
* libgomp.h (enum gomp_map_vars_kind): New.
(gomp_map_vars): Change type of the argument from bool to enum
gomp_map_vars_kind.
* oacc-mem.c (acc_map_data, present_create_copy,
On Wed, Jul 29, 2015 at 10:06:52PM +0300, Ilya Verbin wrote:
@@ -1241,6 +1245,62 @@ GOMP_target_update (int device, const void *unused,
size_t mapnum,
gomp_update (devicep, mapnum, hostaddrs, sizes, kinds, false);
}
+static void
+gomp_exit_data (struct gomp_device_descr *devicep,
On Mon, Jul 06, 2015 at 22:42:10 +0200, Jakub Jelinek wrote:
As has been clarified on omp-lang, we actually shouldn't be mapping or
unmapping the pointer and/or reference, only the array slice itself, except
in target construct (and even for that it is changing from mapping to
private +
On Mon, Jul 06, 2015 at 09:45:30PM +0300, Ilya Verbin wrote:
What exactly do you have in mind here?
void foo (int *p)
{
#pragma omp enter data (to:p[10])
...
#pragma omp exit data (from:p[10])
}
where the latter will only deallocate p[0] ... p[9], but not p?
I've asked for
On Mon, Jul 06, 2015 at 06:34:25PM +0300, Ilya Verbin wrote:
On Thu, Jul 02, 2015 at 00:06:58 +0300, Ilya Verbin wrote:
On Tue, Jun 30, 2015 at 18:10:44 +0200, Jakub Jelinek wrote:
The thing is whether it is actually a good idea to allocate the enter data
allocated objects together.
In
On Mon, Jul 06, 2015 at 19:25:09 +0200, Jakub Jelinek wrote:
On Mon, Jul 06, 2015 at 06:34:25PM +0300, Ilya Verbin wrote:
On Thu, Jul 02, 2015 at 00:06:58 +0300, Ilya Verbin wrote:
The patch is not ready though, I don't know how to unmap GOMP_MAP_POINTER
vars.
In gomp_unmap_vars
On Thu, Jul 02, 2015 at 00:06:58 +0300, Ilya Verbin wrote:
On Tue, Jun 30, 2015 at 18:10:44 +0200, Jakub Jelinek wrote:
The thing is whether it is actually a good idea to allocate the enter data
allocated objects together.
In OpenMP 4.0, generally objects would be allocated and deallocated
On Tue, Jun 30, 2015 at 18:10:44 +0200, Jakub Jelinek wrote:
The thing is whether it is actually a good idea to allocate the enter data
allocated objects together.
In OpenMP 4.0, generally objects would be allocated and deallocated at the
same times, except for multiple host threads trying to
On Tue, Jun 30, 2015 at 03:19:30PM +0300, Ilya Verbin wrote:
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -580,10 +581,16 @@ gomp_unmap_vars (struct target_mem_desc *tgt, bool
do_copyfrom)
bool do_unmap = false;
if (k-refcount 1)
k-refcount--;
- else if
Hi!
This patch implements GOMP_target_enter_exit_data in libgomp, also it fixes a
bug in gomp_map_vars_existing.
make check-target-libgomp passed.
However, I am afraid that there may be some hard-to-find issues (like memory
leaks) in cases of mixed (structured+unstructured) data mappings...
OK
On Tue, Jun 30, 2015 at 14:57:02 +0200, Jakub Jelinek wrote:
On Tue, Jun 30, 2015 at 03:19:30PM +0300, Ilya Verbin wrote:
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -580,10 +581,16 @@ gomp_unmap_vars (struct target_mem_desc *tgt, bool
do_copyfrom)
bool do_unmap = false;
On Tue, Jun 30, 2015 at 06:42:01PM +0300, Ilya Verbin wrote:
On Tue, Jun 30, 2015 at 14:57:02 +0200, Jakub Jelinek wrote:
On Tue, Jun 30, 2015 at 03:19:30PM +0300, Ilya Verbin wrote:
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -580,10 +581,16 @@ gomp_unmap_vars (struct
18 matches
Mail list logo
