If we decided to delete a mapping entry we did not act on it right away
but first issued and waited for memory copies. In the meantime some
other thread might reuse the entry. While there was some logic to avoid
colliding on the actual "deletion" part, there were two races happening:
1) The data transfer back of the thread deleting the entry and
the data transfer back of the thread taking over the entry raced.
2) The update to the shadow map happened regardless if the entry was
actually reused by another thread which left the shadow map in a
inconsistent state.
To fix both issues we will now update the shadow map and delete the
entry only if we are sure the thread is responsible for deletion, hence
no other thread took over the entry and reused it. We also wait for a
potential former data transfer from the device to finish before we issue
another one that would race with it.
Fixes https://github.com/llvm/llvm-project/issues/54216
Differential Revision: https://reviews.llvm.org/D121058
Inline assembly is scary but we need to support it for the OpenMP GPU
device runtime. The new assumption expresses the fact that it may not
have call semantics, that is, it will not call another function but
simply perform an operation or side-effect. This is important for
reachability in the presence of inline assembly.
Differential Revision: https://reviews.llvm.org/D109986
As we mentioned in the code comments for function `ResourcePoolTy::release`,
at some point there could be two identical resources on the two sides of `Next`
mark. It is usually not an issue, unless the following case:
1. Some resources are not returned.
2. We need to iterate the pool and free the element.
That will cause double free, which is the case for event pool. Since we don't release
events hold by the data map, it can happen that the `Next` mark is not reset, and
we have two identical items in the pool. When the pool is destroyed, we will call
`cuEventDestroy` twice on the same event. In the best case, we can only observe
CUDA errors. In the worst case, it can cause internal failures in CUDART and further
crash.
This patch fixes the issue by tracking all resources that have been given using
an `unordered_set`. We don't remove it when a resource is returned. When the pool
is destroyed, we merge the pool (a `vector`) and the set. In this way, we can make
sure that the set contains all resources allocated from the device. We just need
to iterate the set and free the resource accordingly.
For now, only event pool is set to use it. Stream pool is not because we can make
sure all streams are returned when the plugin is destroyed.
Someone might be wondering, why don't we release all events hold in the data map.
That is because, plugins are determined to be destroyed *before* `libomptarget`.
If we can somehow make the plugin outlast `libomptarget`, life will be much
easier.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D122014
This patch adds the necessary AMDGPU calling convention to the ctor /
dtor kernels. These are fundamentally device kenels called by the host
on image load. Without this calling convention information the AMDGPU
plugin is unable to identify them.
Depends on D122504
Fixes#54091
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D122515
This patch solves two problems with the `HostDataToTargetMap` (HDTT
map) which caused races and crashes before:
1) Any access to the HDTT map needs to be exclusive access. This was not
the case for the "dump table" traversals that could collide with
updates by other threads. The new `Accessor` and `ProtectedObject`
wrappers will ensure we have a hard time introducing similar races in
the future. Note that we could allow multiple concurrent
read-accesses but that feature can be added to the `Accessor` API
later.
2) The elements of the HDTT map were `HostDataToTargetTy` objects which
meant that they could be copied/moved/deleted as the map was changed.
However, we sometimes kept pointers to these elements around after we
gave up the map lock which caused potential races again. The new
indirection through `HostDataToTargetMapKeyTy` will allows us to
modify the map while keeping the (interesting part of the) entries
valid. To offset potential cost we duplicate the ordering key of the
entry which avoids an additional indirect lookup.
We should replace more objects with "protected objects" as we go.
Differential Revision: https://reviews.llvm.org/D121057
The unroll pragma did not properly work as the loop bound was not known
when we optimize the runtime and we then added a "unroll disable"
metadata which prevented unrolling later when the bounds were known.
For now we manually unroll to make sure up to 16 elements are handled
nicely. This helps optimizations to look through the argument passing.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D109164
Currently we set ccontext everywhere accordingly, but that causes many
unnecessary function calls. For example, in the resource pool, if we need to
resize the pool, we need to get from allocator. Each call to allocate sets the
current context once, which is unnecessary. In this patch, we set the context
only in the entry interface functions, if needed. Actually in the best way this
should be implemented via RAII, but since `cuCtxSetCurrent` could return error,
and we don't use exception, we can't stop the execution if RAII fails.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D121322
This patch fixes the issue introduced in 14de0820e8 and D120089, that
if dynamic libraries are used, the `CUmodule` array could be overwritten.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D121308
The modules vector was for some reason special which could lead to it
not being of the same size (=num devices). Easiest solution is to treat
it like we do all the other vectors.
An event pool, similar to the stream pool, needs to be kept per device.
For one, events are associated with cuda contexts which means we cannot
destroy the former after the latter. Also, CUDA documentation states
streams and events need to be associated with the same context, which
we did not ensure at all.
Differential Revision: https://reviews.llvm.org/D120142
There are two problems this patch tries to address:
1) We currently free resources in a random order wrt. plugin and
libomptarget destruction. This patch should ensure the CUDA plugin
is less fragile if something during the deinitialization goes wrong.
2) We need to support (hard) pause runtime calls eventually. This patch
allows us to free all associated resources, though we cannot
reinitialize the device yet.
Follow up patch will associate one event pool per device/context.
Differential Revision: https://reviews.llvm.org/D120089
`LIBOMPTARGET_LLVM_INCLUDE_DIRS` is currently checked and included for
multiple times redundantly. This patch is simply a clean up.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D121055
Libomptarget uses some shared variables to track certain internal stated
in the runtime. This causes problems when we have code that contains no
OpenMP kernels. These variables are normally initialized upon kernel
entry, but if there are no kernels we will see no initialization.
Currently we load the runtime into each source file when not running in
LTO mode, so these variables will be erroneously considered undefined or
dead and removed, causing miscompiles. This patch temporarily works
around the most obvious case, but others still exhibit this problem. We
will need to fix this more soundly later.
Fixes#54208.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D121007
When using asynchronous plugin calls, shadow pointer restore could happen before the D2H copy for the entire struct has completed, effectively leaving a device pointer in a host struct.
This patch fixes the problem by delaying restore's to after a synchronization happens (target regions) and by calling early synchronization (target update).
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119968
The runtime uses thread state values to indicate when we use an ICV or
are in nested parallelism. This is done for OpenMP correctness, but it
not needed in the majority of cases. The new flag added is
`-fopenmp-assume-no-thread-state`.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D120106
`bug49334.cpp` has one issue that causes flaky result reported in #53730.
The root cause is `BlockedC` is never initialized but in `BlockMatMul_TargetNowait`
it is directly read and written (via `+=`). Fixes#53730.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D119988
The `IsSPMD` global can only be read by threads other than the main
thread *after* initialization is complete. To allow usage of
`mapping::getBlockSize` before initialization is done, we can pass the
`IsSPMD` state explicitly. This is similar to other APIs that take
`IsSPMD` explicitly to avoid such a race, e.g.,
`mapping::isInitialThreadInLevel0(IsSPMD)`
Fixes https://github.com/llvm/llvm-project/issues/53857
This patch adds a new target to the OpenMP CPU offloading tests. This
tests the usage of the new driver for CPU offloading. If this all works
then we can move to transition to the new driver as the default.
Depends on D119613
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119736
Currently whenever we compile the device runtime we get the following
'Mapping.cpp:32:32: warning: inline function '_OMP::impl::getGridValue'
is not defined [-Wundefined-inline]' warning. This can be silenced by
removing the constexpr attribute for this function. Doing this doesn't
change the generated bitcode at all but prevents the screen from getting
filled with warnings whenver we build the runtime.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119747
This patch fixes the issue that the for loop in `applyToShadowMapEntries`
is infinite because `Itr` is not incremented in `CB`. Fixes#53727.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119471
`bug49334.cpp` directly uses `!=` to compare two floating point values,
which is almost wrong.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D119485
Currently we have a hard team limit, which is set to 65536. It says no matter whether the device can support more teams, or users set more teams, as long as it is larger than that hard limit, the final number to launch the kernel will always be that hard limit. It is way less than the actual hardware limit. For example, my workstation has GTX2080, and the hardware limit of grid size is 2147483647, which is exactly the largest number a `int32_t` can represent. There is no limitation mentioned in the spec. This patch simply removes it.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119313
This patch refines the logic to determine grid size as previous method
can escape the check of whether `CudaBlocksPerGrid` could be greater than the actual
hardware limit.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119311
The 'bug49779.cpp' test has been failing recently. This is because the
runtime is sufficiently complex when using nested parallelism without
optimizations that the CUDA tools cannot statically determine the stack
size. Because of this the kernel can exceed the thread stack size and
crash. Work around this using the 'LIBOMPTARGET_STACK_SIZE' environment
variable and add an FAQ entry for this situation.
Fixes#53670
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D119357
This patch manually adds the runtime include files to the list of
dependencies when we build the bitcode runtime library. Previously if
only the header was changed we would not recompile the source files.
The solution used here isn't optimal because every source file not has a
dependency on each header file regardless of if it was actually used by
that file.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D119254
This patch enables running the new driver tests for AMDGPU. Previously
this was disabled because some tests failed. This was only because the
new driver tests hadn't been listed as unsupported or expected to fail.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D119240
This patch replaces the ValueRAII pointer with a default 'nullptr'
value. Previously this was initialized as a reference to an existing
variable. The use of this variable caused overhead as the compiler could
not look through the uses and determine that it was unused if 'Active'
was not set. Because of this accesses to the variable would be left in
the runtime once compiled.
Fixes#53641
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D119187
This patch completely removes the old OpenMP device runtime. Previously,
the old runtime had the prefix `libomptarget-new-` and the old runtime
was simply called `libomptarget-`. This patch makes the formerly new
runtime the only runtime available. The entire project has been deleted,
and all references to the `libomptarget-new` runtime has been replaced
with `libomptarget-`.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D118934
Due to num_threads (probably also other reasons) we cannot assume
explicit barriers are always executed by all threads in an aligned
fashion. We can optimize them if that property can be proven but
that is different.
Ron Lieberman