From: Matthew Malcomson <[email protected]>
I'd previously split these patches up into logically independent
changes, but since patches have been written on top of the others have
just made maintainers jobs more difficult.
- Sebastian just pointed out that I'd included the wrong link in my
latest email so the combination was incorrect, so I'll go the way less
likely to make mistakes and send everything as a single patch series.
Hence sending up the patch series as a complete patchset ordered on top
of each other. In order to do that I rebased the "Move thread task
re-initialisation into threads" patch on top of the others (some order
had to be chosen since neither order cleanly applies). This is the
order in which I did most of my testing (and TSAN testing was done on
the combination of all patches in the order sent here).
Apologies for the noise & extra back-and-forth around attempting to
apply patches.
Including original cover letter for the patchset below here:
------------------------------
Cc'ing in maintainers of nvptx, gcn, and rtems ports for target specific
changes (especially with request for runtime testing).
After having updated the target code, looked into various TODO's and ran
more testing I've combined my previous patches into one patchset. This
patchset drastically improves the performance on the micro-benchmark
119588.
This micro-benchmark represents a significant slowdown in some OMP uses
in NVPL BLAS running GEMM routines on small matrices with a high level
of parallelism. (High level of parallelism due to other routines in the
code benefiting from many threads, and there being no low-overhead way
to change the level of parallelism between routines).
This patchset has 5 commits:
1) Is a fix for PR122314. It ensures that GOMP tasks are executed
logically in the region where they are scheduled.
2) Is a fix for PR122356. It ensures there is a memory synchronisation
point between tasks being run in the barrier and the barrier
continuing.
3) Changes the linux/ barrier implementation from the "centralized"
method currently used to a combination of a "linear" barrier gather
and "centralized" barrier release.
- I see this gives about a 3x improvement on time through a highly
contended barrier on a 144 core machine.
4) Follows the LLVM example and "wait" between parallel regions *inside*
the barrier rather than between two barriers. This halves the
overhead from barriers on many consecutive parallel regions.
5) Reduces the data structure initialisation overhead when starting a
new parallel region. Rather than have the primary thread initialise
each threads data while each secondary thread is waiting the primary
thread stores common data and lets each secondary thread initialise
its thread-specific data from that shared information.
Patches (1), (2), and (5) could all be made independent (with some
adjustment for patch context). Patch (4) requires patch (3) and patch
(3) introduces some less-pleasant code structure that the changes in
patch (4) help fix.
In order to use the feature introduced in patch (3) we have to change
the barrier API to pass an ID. For patch (3) alone we also need to
introduce some relatively awkward interfaces for adjusting the size of
the barrier.
Patch (4) removes that need for the new awkward interface (the only
barrier that needs size adjustment is now no longer in the fast path).
Since I hope to have both patches in I have only made changes for other
targets to build on top of patch (4). This in order to avoid writing
the implementation for this awkward interface that I intend to never be
actually used.
N.b. when I did bootstrap & regtest on the posix/ target I saw flaky
tests before and after. Believe the same flaky tests.
Matthew Malcomson (5):
libgomp: Enforce tasks executed lexically after scheduled
libgomp: Ensure memory sync after performing tasks
libgomp: Implement "flat" barrier for linux/ target
libgomp: Removing one barrier in non-nested thread loop
libgomp: Move thread task re-initialisation into threads
libgomp/barrier.c | 4 +-
libgomp/config/gcn/bar.c | 53 +-
libgomp/config/gcn/bar.h | 98 ++-
libgomp/config/gcn/team.c | 2 +-
libgomp/config/linux/bar.c | 798 ++++++++++++++++--
libgomp/config/linux/bar.h | 331 +++++++-
libgomp/config/linux/futex_waitv.h | 129 +++
libgomp/config/linux/simple-bar.h | 66 ++
libgomp/config/linux/wait.h | 15 +-
libgomp/config/nvptx/bar.c | 36 +-
libgomp/config/nvptx/bar.h | 89 +-
libgomp/config/nvptx/team.c | 2 +-
libgomp/config/posix/bar.c | 41 +-
libgomp/config/posix/bar.h | 93 +-
libgomp/config/posix/pool.h | 1 +
libgomp/config/posix/simple-bar.h | 10 +-
libgomp/config/rtems/bar.c | 185 +++-
libgomp/config/rtems/bar.h | 97 ++-
libgomp/libgomp.h | 21 +-
libgomp/single.c | 4 +-
libgomp/task.c | 66 +-
libgomp/team.c | 292 ++++++-
.../testsuite/libgomp.c++/task-reduction-20.C | 136 +++
.../testsuite/libgomp.c++/task-reduction-21.C | 140 +++
libgomp/testsuite/libgomp.c/pr122314.c | 36 +
libgomp/testsuite/libgomp.c/pr122356.c | 33 +
.../libgomp.c/primary-thread-tasking.c | 80 ++
libgomp/work.c | 26 +-
28 files changed, 2614 insertions(+), 270 deletions(-)
create mode 100644 libgomp/config/linux/futex_waitv.h
create mode 100644 libgomp/config/linux/simple-bar.h
create mode 100644 libgomp/testsuite/libgomp.c++/task-reduction-20.C
create mode 100644 libgomp/testsuite/libgomp.c++/task-reduction-21.C
create mode 100644 libgomp/testsuite/libgomp.c/pr122314.c
create mode 100644 libgomp/testsuite/libgomp.c/pr122356.c
create mode 100644 libgomp/testsuite/libgomp.c/primary-thread-tasking.c
--
2.43.0
