Summary:
Previously we had this `LIBOMPTARGET_ENABLED` variable which controlled
including `libomptarget`. This is now redundant since it's controlled by
`LLVM_ENABLE_RUNTIMES`. However, this had the extra effect of not
building it when given unsupported targets. THis was lost during the
move to `offload`. This patch moves this logic back and makes the
`offload` target just quit without doing anything if used on an
unsupported architecture.
https://github.com/llvm/llvm-project/issues/91881https://github.com/llvm/llvm-project/issues/91819
---------
Co-authored-by: Sylvestre Ledru <sylvestre@debian.org>
Allow non-constants in the `sizes` clause such as
```
#pragma omp tile sizes(a)
for (int i = 0; i < n; ++i)
```
This is permitted since tile was introduced in [OpenMP
5.1](https://www.openmp.org/spec-html/5.1/openmpsu53.html#x78-860002.11.9).
It is possible to sneak-in negative numbers at runtime as in
```
int a = -1;
#pragma omp tile sizes(a)
```
Even though it is not well-formed, it should still result in every loop
iteration to be executed exactly once, an invariant of the tile
construct that we should ensure. `ParseOpenMPExprListClause` is
extracted-out to be reused by the `permutation` clause of the
`interchange` construct. Some care was put into ensuring correct behavior
in template contexts.
AIX has the `/proc` filesystem where `/proc/<pid>/lwp/<tid>/lwpsinfo` has
the thread state in binary, similar to Linux's
`/proc/<pid>/task/<tid>/stat` where the state is in ASCII. However, the
definition of state info `R` in `lwpsinfo` is `runnable`. In Linux,
state `R` means the thread is `running`. Therefore, `lwpsinfo` is not
ideal for our purpose of getting the current load of the system. This
patch uses `perfstat_cpu()` in AIX system library `libperfstat.a` to
obtain the number of threads current running on logical CPUs.
Root cause: Segmentation fault is caused by null pointer dereference
inside the __kmpc_fork_call_if function at
https://github.com/llvm/llvm-project/blob/main/openmp/runtime/src/z_Linux_asm.S#L1186
. __kmpc_fork_call_if is missing case to handle argc=0 .
Fix: Added a check inside the __kmp_invoke_microtask function to handle
the case when argc is 0.
---------
Co-authored-by: Singh <chasingh@amd.com>
When a child process is forked with OpenMP already initialized, the
child process resets its affinity mask and sets proc-bind-var to false
so that the entire original affinity mask is used. This patch corrects
an issue with the affinity initialization code setting affinity to
compact instead of none for this special case of forked children.
The test trying to catch this only testing explicit setting of
KMP_AFFINITY=none. Add test run for no KMP_AFFINITY setting.
Fixes: #91098
* Serial teams now use a stack (similar to dispatch buffers)
* Serial teams always use `t_task_team[0]` as the task team and the
second pointer is a next pointer for the stack
`t_task_team[1]` is interpreted as a stack of task teams where each
level is a nested level
```
inner serial team outer serial team
[ t_task_team[0] ] -> (task_team) [ t_task_team[0] ] -> (task_team)
[ next ] ----------------> [ next ] -> ...
```
* Remove the task state memo stack from thread structure.
* Instead of a thread-private stack, use team structure to store
th_task_state of the primary thread. When coming out of a parallel,
restore the primary thread's task state. The new field in the team
structure doesn't cause sizeof(team) to change and is in the cache line
which is only read/written by the primary thread.
Fixes: #50602Fixes: #69368Fixes: #69733Fixes: #79416
Revert the portion of https://github.com/llvm/llvm-project/pull/75125
which modified the LIBOMP_HEADERS_INSTALL_PATH in standalone build.
This change is harmful for real standalone builds (i.e. builds where we
build openmp by itself), since it tries to overwrite the `omp.h` inside
the build compiler. For all-in-one builds with clang, testing shows this
change is unnecessary as https://github.com/llvm/llvm-project/pull/88007
already set up that build configuration so that omp.h will be put into
the project build's `clang` resource directory.
This patch implements `__kmp_is_address_mapped()` for AIX by calling
`loadquery()` to get the load info of the process and then checking if
the address falls within the range of the data segment of one of the
loaded modules.
This patch changes to use system call `syssmt()` instead of
`lpar_get_info()` to get the number of SMTs (logical processors) per
physical processor for AIX. `lpar_get_info()` gives the max number of
SMTs that the physical processor can support while `syssmt()` returns
the number that is currently configured.
In a nutshell, this moves our libomptarget code to populate the offload
subproject.
With this commit, users need to enable the new LLVM/Offload subproject
as a runtime in their cmake configuration.
No further changes are expected for downstream code.
Tests and other components still depend on OpenMP and have also not been
renamed. The results below are for a build in which OpenMP and Offload
are enabled runtimes. In addition to the pure `git mv`, we needed to
adjust some CMake files. Nothing is intended to change semantics.
```
ninja check-offload
```
Works with the X86 and AMDGPU offload tests
```
ninja check-openmp
```
Still works but doesn't build offload tests anymore.
```
ls install/lib
```
Shows all expected libraries, incl.
- `libomptarget.devicertl.a`
- `libomptarget-nvptx-sm_90.bc`
- `libomptarget.rtl.amdgpu.so` -> `libomptarget.rtl.amdgpu.so.18git`
- `libomptarget.so` -> `libomptarget.so.18git`
Fixes: https://github.com/llvm/llvm-project/issues/75124
---------
Co-authored-by: Saiyedul Islam <Saiyedul.Islam@amd.com>
The new collapse test cases define `MAX_THREADS` to be 256 and use all
available threads/logical processors on the system. This triples the
testing time on an AIX machine that has 128 logical processors. This
patch changes to use half of available logical processors to avoid over
subscribing because there are other libomp tests running at the same
time, including 2 other such collapse tests.
Summary:
Previously, the R&R support was global state initialized by a global
constructor. This is bad because it prevents us from adequately
constraining the lifetime of the library. Additionally, we want to
minimize the amount of global state floating around.
This patch moves the R&R support into a plugin member like everything
else. This means there will be multiple copies of the R&R implementation
floating around, but this was already the case given the fact that we
currently handle everything with dynamic libraries.
This patch uses a memory fence in function `__kmp_dispatch_next()` to
flush pending memory write invalidates before incrementing the
`volatile` variable `buffer_index` to fix intermittent time-outs of
OpenMP runtime LIT test cases `env/kmp_set_dispatch_buf.c` and
`worksharing/for/kmp_set_dispatch_buf.c`, noting that the same is needed
for incrementing `buffer_index` in function `__kmpc_next_section()`
(line 2600 of `kmp_dispatch.cpp`).
Summary:
The previous patch mistakenly merged these when they indeed need to be
treated like separate triples because it's what's passed to the test
suite.
This was initially reported here (including stacktraces):
https://stackoverflow.com/questions/78183545/does-compiling-imagick-with-openmp-enabled-in-freebsd-13-2-cause-sched-yield
If `__kmp_register_library_startup()` detects that another instance of
the library is present, `__kmp_is_address_mapped()` is eventually
called. which uses `kmpc_alloc()` to allocate memory. This function
calls `__kmp_entry_thread()` to access the thread-local memory pool,
which is a bad idea during initialization. This macro internally calls
`__kmp_get_global_thread_id_reg()` which sets the bootstrap lock at the
beginning (before calling `__kmp_register_library_startup()`).
The fix is to use `KMP_INTERNAL_MALLOC()`/`KMP_INTERNAL_FREE()` instead
of `kmpc_malloc()`/`kmpc_free()`. `KMP_INTERNAL_MALLOC` and
`KMP_INTERNAL_FREE` do not use any bootstrap locks. They just translate
to `malloc()`/`free()` and are meant to be used during library
initialization before other library-specific allocators have been
initialized.
Fixes: #86684
This patch fixes the test config so that it works for
`tasking/omp50_taskdep_depobj.c` which uses different flags to test with
compiler's `omp.h`.
* set test environment variable `OBJECT_MODE` to `64` if it is set
explicitly to `64` in the AIX environment. `OBJECT_MODE` is default to
`32` and is recognized by AIX compilers and toolchain. In this way, we
don't need to set `-m64` for all compiler flags for 64-bit mode
* add option `-Wl,-bmaxdata` to 32-bit `test_openmp_flags` used by
`tasking/omp50_taskdep_depobj.c`
Summary:
Relanding after reverting, only applies to AMDGPU for now.
This patch adds an implementation of printf that's provided by the GPU
C library runtime. This pritnf currently implemented using the same
wrapper handling that OpenMP sets up. This will be removed once we have
proper varargs support.
This printf differs from the one CUDA offers in that it is synchronous
and uses a finite size. Additionally we support pretty much every
format specifier except the %n option.
Depends on #85331
Summary:
These headers are a part of the compiler's resource directory once
installed. However, they are currently placed in the binary directory
temporarily. This makes it more difficult to use the compiler out of the
build directory and will cause issues when moving to `liboffload`. This
patch changes the logic to write these instead to the copmiler's
resource directory inside of the build tree.
NOTE: This doesn't change the Fortran headers, I don't know enough about
those and it won't use the same directory.
…irectory (#88007)"
This reverts commit 8671429151d5e67d3f21a737809953ae8bdfbfde.
This commit broke the flang build, so I'm reverting it. See the comments
in merge request #88007 for more information.
Summary:
These headers are a part of the compiler's resource directory once
installed. However, they are currently placed in the binary directory
temporarily. This makes it more difficult to use the compiler out of the
build directory and will cause issues when moving to `liboffload`. This
patch changes the logic to write these instead to the copmiler's
resource directory inside of the build tree.
NOTE: This doesn't change the Fortran headers, I don't know enough about
those and it won't use the same directory.
Users can put a : in front of KMP_HW_SUBSET to indicate that the
specified subset is an "absolute" subset. Currently, when a user puts
KMP_HW_SUBSET=1t. This gets translated to KMP_HW_SUBSET="*s,*c,1t",
where * means "use all of". If a user wants only one thread as the
entire topology they can now do KMP_HW_SUBSET=:1t.
Along with the absolute syntax is a fix for newer machines and making
them easier to use with only the 3-level topology syntax. When a user
puts KMP_HW_SUBSET=1s,4c,2t on a machine which actually has 4 layers,
(say 1s,2m,3c,2t as the entire machine) the user gets an unexpected "too
many resources asked" message because KMP_HW_SUBSET currently translates
the "4c" value to mean 4 cores per module. To help users out, the
runtime can assume that these newer layers, module in this case, should
be ignored if they are not specified, but the topology should always
take into account the sockets, cores, and threads layers.
Summary:
This patch adds an implementation of `printf` that's provided by the GPU
C library runtime. This `pritnf` currently implemented using the same
wrapper handling that OpenMP sets up. This will be removed once we have
proper varargs support.
This `printf` differs from the one CUDA offers in that it is synchronous
and uses a finite size. Additionally we support pretty much every format
specifier except the `%n` option.
Depends on https://github.com/llvm/llvm-project/pull/85331
When a nested parallel region ends, the runtime calls __kmp_join_call().
During this call, the primary thread of the nested parallel region will
reset its tid (retval of omp_get_thread_num()) to what it was in the
outer parallel region. A data race occurs with the current code when
another worker thread from the nested inner parallel region tries to
steal tasks from the primary thread's task deque. The worker thread
reads the tid value directly from the primary thread's data structure
and may read the wrong value.
This change just uses the calculated victim_tid from execute_tasks()
directly in the steal_task() routine rather than reading tid from the
data structure.
Fixes: #87307
`hw.logicalcpu` returns the available logical core count
Fix build error for watchOS
```
runtime/src/z_Linux_util.cpp:1821:8: error: 'host_info' is unavailable: not available on watchOS
rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
^
/Applications/Xcode_15.2.app/Contents/Developer/Platforms/WatchOS.platform/Developer/SDKs/WatchOS10.2.sdk/usr/include/mach/mach_host.h:82:15: note: 'host_info' has been explicitly marked unavailable here
kern_return_t host_info
^
1 warning and 1 error generated.
make[2]: *** [runtime/src/CMakeFiles/omp.dir/z_Linux_util.cpp.o] Error 1
```
detect `aarch64_32` with compiler defined macro `__ARM64_ARCH_8_32__`
reuse ARM `__kmp_unnamed_critical_addr` and add `KMP_PREFIX_UNDERSCORE`
macro like AARCH64
reuse AARCH64 `__kmp_invoke_microtask`
build log for watchos armv7k + arm64_32 and watchos simulator x86_64 +
arm64
https://github.com/nihui/action-protobuf/actions/runs/8520684611/job/23337305030
Summary:
The previous code would potentially make it smaller if a device with a
lower ID touched it later. Also we should minimize changes to the state
for multi threaded reasons. This just sets up an owned slot for each at
initialization time.
The hidden helper team pre-allocates the gtid space [1,
num_hidden_helpers] (inclusive). If regular host threads are allocated,
then put back in the thread pool, then the hidden helper team is
initialized, the hidden helper team tries to allocate the threads from
the thread pool with gtids higher than [1, num_hidden_helpers]. Instead,
have the hidden helper team fork OS threads so the correct gtid range
used for hidden helper threads.
Fixes: #87117
Similar to H2D and D2H, use synchronous mode for large data transfers
beyond a certain size for D2D as well. As with H2D and D2H, this size is
controlled by an env-var.
Summary:
The current implementation of RPC tied everything to device IDs and
forced us to do init / shutdown to manage some global state. This turned
out to be a bad idea in situations where we want to track multiple
hetergeneous devices that may report the same device ID in the same
process.
This patch changes the interface to instead create an opaque handle to
the internal device and simply allocates it via `new`. The user will
then take this device and store it to interface with the attached
device. This interface puts the burden of tracking the device identifier
to mapped d evices onto the user, but in return heavily simplifies the
implementation.
Summary:
This patch removes most of the special handling from the
`PluginAdaptorTy` in preparation for changing this to be the
`GenericPluginTy`. Doing this requires that the OpenMP specific handling
of stuff like device offsets be contained within the OpenMP plugin
manager. Generally this was uninvasive expect for the change to tracking
the offset and size of the used devices. The eaiest way I could think to
do this was to use some maps, which double as indicators for which
plugins have devices active. This should not affect the logic.
The __kmpc_omp_taskwait_deps_51 allocates a kmp_depnode_t node on its
stack, and there is currently a race condition where another thread
might still be accessing that node after the function has returned and
its stack frame was released.
While the function does wait until the node's npredecessors count has
reached zero before exiting, there is still a window where the function
that last decremented the npredecessors count assumes the node is still
accessible.
For heap-allocated kmp_depnode_t nodes, this normally works via a
separate ndeps count that only reaches zero at the point where no
accesses to the node are expected at all; in fact, at this point the
heap allocation will be freed.
For this case of a stack-allocated kmp_depnode_t node, it therefore
makes sense to similarly respect the ndeps count; we need to wait until
this reaches 1 (not 0, because it is not heap-allocated so there's
always one extra count to prevent it from being freed), before we can
safely deallocate our stack frame.
As this is expected to be a short race window of only a few
instructions, it should be fine to just use a busy wait loop checking
the ndeps count.
Fixes: https://github.com/llvm/llvm-project/issues/85963
Summary:
Previously we had an interface that checked these functions pointers to
see if they are implemented by the plugin. This was removed as currently
every single function is implemented as a part of the common interface.
These checks are now always true and do nothing.
Summary:
The plan is to remove the entire plugin interface and simply use the
`GenericPluginTy` inside of `libomptarget` by statically linking against
it. This means that inside of `libomptarget` we will simply do
`Plugin.data_alloc` without the dynamically loaded interface. To reduce
the amount of code required, this patch simply moves all of the RTL
implementation functions inside of the Generic device. Now the
`__tgt_rtl_` interface is simply a shallow wrapper that will soon go
away. There is some redundancy here, this will be improved later. For
now what is important is minimizing the changes to the API.
Summary:
We have a plugin singleton that implements the Plugin interface. This
then spawns separate device and kernels. Previously when these needed to
reach into the global singleton they would use the `PluginTy::get`
routine to get access to it. In the future we will move away from this
as the lifetime of the plugin will be handled by `libomptarget`
directly. This patch removes uses of this inside of the plugin
implementaion themselves by simply keeping a reference to the plugin
inside of the device.
The external `__tgt_rtl` functions still use the global method, but will
be removed later.
Summary:
This patch factors common functions out of the `Plugin` interface prior
to its removal in a future patch. This simply temporarily renames it to
`PluginTy` so that we could re-use `Plugin::check` internally as this
needs to be defined statically per plugin now. We can refactor this
later.
The future patch will delete `PluginTy` and `PluginTy::get` entirely.
This simply tries to minimize a few changes to make it easier to land.
Summary:
It's possible for the OpenMP offloading plugins to be build before
tablegen is run despite the fact that we rely on it. Simply make it
depend on it currently.
