Summary:
Previously this test hanged indefinitely on NVPTX. This was due to an
issue fixed previously where we would wait indefinitely inside the CUDA
runtime waiting for the kernel to complete if it was blocked on the RPC
server. This patch enables this test again now that it can run without
deadlocking, at least on CUDA 12.2.
Summary:
The RPC server is responsible for providing host services from the GPU.
Generally, the client running on the GPU will spin in place until the
host checks the server. Inside the runtime, we elected to have the user
thread do this checking while it would be otherwise waiting for the
kernel to finish. However, for Nvidia this caused problems when
offloading to a target region that requires a copy back.
This is caused by the implementation of `dataRetrieve` on Nvidia. We
initialize an asynchronous copy-back on the same stream that the kernel
is running on. This creates an implicit sync on the kernel to finish
before we issue the D2H copy, which we then wait on. This implicit sync
happens inside of the CUDA runtime. This is problematic when running the
RPC server because we need someone to check the RPC server. If no one
checks the RPC server then the kernel will never finish, meaning that
the memcpy will never be issued and the program hangs. This patch adds
an explicit check for unfinished work on the stream and waits for it to
complete.
Summary:
Normally, the implementation of `puts` simply writes a second newline
charcter after printing the first string. However, because the GPU does
everything in batches of the SIMT group size, this will end up with very
poor output where you get the strings printed and then 1-64 newline
characters all in a row. Optimizations like to turn `printf` calls into
`puts` so it's a good idea to make this produce the expected output.
The least invasive way I could do this was to add a new opcode. It's a
little bloated, but it avoids an unneccessary and slow send operation to
configure this.
This patch adds initial lowering for DeclareTargetAttr on
GlobalOp's utilising registerTargetGlobalVariable
and getAddrOfDeclareTargetVar from the
OMPIRBuilder.
It also adds initial processing of declare target map
operands, populating the combinedInfo that the
OMPIRBuilder requires to generate kernels and
it's kernel argument structure.
The combination of these additions allows simple mapping
of declare target globals to Target regions, as such a simple
runtime test showcasing this and testing it has been added.
The patch currently does not factor in filtering
based on device_type clauses (e.g. no emission of
globals for device if host specified), this will come in
a future iteration. And for the moment it's only been
tested with 1-D arrays and basic fortran data types,
more complex types (such as user defined derived
types from Fortran, allocatables or Fortran pointers)
may need further work.
reviewers: kiranchandramohan, skatrak
Differential Revision: https://reviews.llvm.org/D149368
VE supports up to 64 threads per a VE process. So, we limit the number
of threads defined by KMP_MAX_NTH. We also modify the __kmp_sys_max_nth
initialization to use KMP_MAX_NTH as a limit.
This will make it easy for callers to see issues with and fix up calls
to createTargetMachine after a future change to the params of
TargetMachine.
This matches other nearby enums.
For downstream users, this should be a fairly straightforward
replacement,
e.g. s/CodeGenOpt::Aggressive/CodeGenOptLevel::Aggressive
or s/CGFT_/CodeGenFileType::
The /tmp fallback for /dev/shm did not write to a fixed filename, so multiple instances of the runtime would not be able to detect each other. Now, we create the /tmp file in much the same way as the /dev/shm file was created, since mkstemp approach would not work to create a file that other instances of the runtime would detect. Also, add the environment variable method as a third fallback to /dev/shm and /tmp for library registration, as some systems do not have either. Also, add ability to fallback to a subsequent method should a failure occur during any part of the registration process. When unregistering, it is assumed that the method chosen during registration should work, so errors at that point are ignored. This also avoids a problem with multiple threads trying to unregister the library.
This commit removes an optimization that skips the initialization of the
reduction struct if the number of threads in a team is 1. This
optimization
caused a bug with Hidden Helper Threads. When the task group is
initially
initialized by the master thread but a Hidden Helper Thread executes a
target
nowait region, it requires the reduction struct initialization to
properly
accumulate the data.
This commit also adds a LIT test for issue #57522 to ensure that the
issue is
properly addressed and that the optimization removal does not introduce
any
regressions.
Fixes: #57522
Fixes: https://github.com/llvm/llvm-project/issues/65104
When a user assigns devices to target regions it may happen that
different identifiers will map onto the same id within different
plugins. This will lead to situations where callbacks will become much
harder to read, as ambiguous identifiers are reported.
We fix this by collecting the index-offset upon general RTL
initialization. Which in turn, allows to calculate the unique,
user-observable device id.
In omptarget.cpp, there is a targetDataMapper which does mapper related
operations.
In interface.cpp, targetDataMapper function template is simply needed
for handling TargetAsyncInfoTy. Thus it is better to name it as simple
as targetData similar to targetKernel.
Change to use VE_LD_LIBRARY_PATH for VE instead of LD_LIBRARY_PATH. The
VE is connected to the host, and compiled test programs for VE is
invoked on the host and transferred to the VE. If programs are compiled
for the host, we use LD_LIBRARY_PATH. Otherwise, we use
VE_LD_LIBRARY_PATH.
Support OpenMP runtime library on VE. This patch makes OpenMP compilable
for VE architecture. Almost all tests run correctly on VE.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D159401
Based on https://en.cppreference.com/w/c/memory/aligned_alloc, the
`size` is supposed
to be a multiple of `alignment`, and it is implementation defined
behavior if not.
We have a non-conformant use in `kmp_barrier.h` when allocating
distribute barrier.
The size of the barrier is 576 and the alignment is `4*CACHE_LINE`,
which is 256
on most systems. Apparently it works perfectly fine for Linux and
Intel-based Mac,
but not for Apple Silicon based Mac.
Fix#63194.
Summary:
The `omp_get_num_procs()` function should return the amount of
parallelism availible. On the GPU, this was not defined. We have elected
to define this function as the maximum amount of wavefronts / warps that
can be simultaneously resident on the device. For AMDGPU this is the
number of CUs multiplied byth CU's per wave. For NVPTX this is the
maximum threads per SM divided by the warp size and multiplied by the
number of SMs.
The function assumes that `__kmp_gtid_get_specific` always returns a valid gtid.
That is not always true, because when creating the key for thread-specific data,
a destructor is assigned. The dtor will be called at thread exit. However, before
the dtor is called, the thread-specific data will be reset to NULL first
(https://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_key_create.html):
> At thread exit, if a key value has a non-NULL destructor pointer, and the thread
> has a non-NULL value associated with that key, the value of the key is set to NULL.
This will lead to that `__kmp_gtid_get_specific` returns `KMP_GTID_DNE`.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D159369
The outlined function is typically invoked by using
`__kmp_invoke_microtask`,
which is written in asm. D138495 introduces a new interface function for
parallel
region for OpenMPIRBuilder, where the outlined function is called via
the function
pointer. For some reason, it works perfectly well on x86 and x86-64
system, but
doesn't work on Apple Silicon. The 3rd argument in the callee is always
`nullptr`, even
if it is not in caller. It appears `x2` always contains `0x0`. This
patch adopts
the typical method to invoke the function pointer. It works on my M2
Ultra Mac.
Fix#63194.
This patch ensures that the locally built version of flang when building in-tree. `find_program` sometimes used the wrong executable if a different copy of flang was installed.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D159161
Previously, the test ran a section with
#pragma omp target thread_limit(4)
and expected it to execute exactly 4 times, even though it would
in practice execute min(cores, 4) times.
Increment a counter and check that it executed 1-4 times.
Differential Revision: https://reviews.llvm.org/D159311
This patch adds a test that uses a target region to set a scalar value. It also
adds rules in lit.cfg to handle fortran testing.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D159216
In `__kmp_get_global_thread_id`, if the gtid mode is 1, after getting the gtid
from TLS, it will store the gtid value to the thread stack maintained in the thread
descriptor. However, `__kmp_get_global_thread_id` can be called when the library
is destructed, after the corresponding thread info has been release. This will
cause a segment fault. This can happen on an Intel-based Mac.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D159324
This patch fixes the flang detection in the openmp fortran offloading test.
Reviewed By: jsjodin
Differential Revision: https://reviews.llvm.org/D158546
Update DeviceRTL and the AMDGPU plugin to support code
object version 5. Default is code object version 4.
CodeGen for __builtin_amdgpu_workgroup_size generates code
for cov4 as well as cov5 if -mcode-object-version=none
is specified. DeviceRTL compilation passes this argument
via Xclang option to generate abi-agnostic code.
Generated code for the above builtin uses a clang
control constant "llvm.amdgcn.abi.version" to branch on
the abi version, which is available during linking of
user's OpenMP code. Load of this constant gets eliminated
during linking.
AMDGPU plugin queries the ELF for code object version
and then prepares various implicitargs accordingly.
Differential Revision: https://reviews.llvm.org/D139730
Reviewed By: jhuber6, yaxunl
As discussed on the weekly OpenMP meeting on the second of August 2023, the default version
in the OpenMP documentation shoud be changed from OpenMP 5.0 to 5.1.
Differential Revision: https://reviews.llvm.org/D156901
At the moment Archer segfaults due to a null-pointer access, if an application
uses taskwait with depend clause as used in the two new tests.
This patch cleans up the task_schedule function, moves semantic blocks into
functions and replaces the if blocks by a single switch statement. The switch
statement will warn, when new enum values are added in OMPT and makes clear
what code is executed for the different cases.
With free-agent tasks coming up in OpenMP 6.0, we should expect more
null-pointer task_data, so additional null-pointer checks were added.
We also cannot rely on having an implicit task on the stack, so the
BarrierIndex is stored during task creation.
Differential Revision: https://reviews.llvm.org/D158072
Since td_allow_completion_event is a member of the taskdata struct, not all
firstprivate/shared variables are stored at the end of the task memory
allocation. Simply report the whole allocation instead.
Furthermore, the function should always return 0 since in no case there is
another block to report.
Differential Review: https://reviews.llvm.org/D158080
offloading
- This patch adds support for thread_limit clause on target directive according to OpenMP 51 [2.14.5]
- The idea is to create an outer task for target region, when there is a thread_limit clause, and manipulate the thread_limit of task instead. This way, thread_limit will be applied to all the relevant constructs enclosed by the target region.
Differential Revision: https://reviews.llvm.org/D152054
D152014 introduced an optimization that favors more smaller blocks over
fewer larger blocks, even if user sets `thread_limit` explicitly. This patch changes
the behavior to honor user value.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D158802
The changes in D157738 allowed for us to emit stub globals on the device
in the offloading entry section. These globals contain the addresses of
device functions and allow us to map host functions to their
corresponding device equivalent. This patch provides the initial support
required to build a table on the device to lookup the associated value.
This is done by finding these entries and creating a global table on the
device that can be searched with a simple binary search.
This requires an allocation, which supposedly should be automatically
freed at plugin shutdown. This includes a basic test which looks up device
pointers via a host pointer using the added function. This will need to be built
upon to provide full support for these calls in the runtime.
To support reverse offloading it would also be useful to provide a reverse table
that allows us to get host functions from device stubs.
Depends on D157738
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D157918
Before, we checked the parallel region only once, and ignored updates in
the KernelInfo for the parallel region that happened later. This caused
us to think nested parallel sections are not present even if they are,
among other things.
Make the generic-plugin report a corresponding CU kind -- instead of 'unknown'.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D158542
