Eachempati.
This patch adds clang (parsing, sema, serialization, codegen) support for the 'depend' clause on the 'taskwait' directive.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D113540
Teach the HWLOC topology method how to detect Atom and Core
types so hybrid CPUs are properly detected and represented when using
the HWLOC topology method.
Differential Revision: https://reviews.llvm.org/D112270
The current implementation of Windows Processor Groups has
a separate topology method to handle them. This patch deprecates
that specific method and uses the regular CPUID topology
method by default and inserts the Windows Processor Group objects
in the topology manually.
Notes:
* The preference for processor groups is lowered to a value less than
socket so that the user will see sockets in the KMP_AFFINITY=verbose
output instead of processor groups when sockets=processor groups.
* The topology's capacity is modified to handle additional topology layers
without the need for reallocation.
* If a user asks for a granularity setting that is "above" the processor
group layer, then the granularity is adjusted "down" to the processor
group since this is the coarsest layer available for threads.
Differential Revision: https://reviews.llvm.org/D112273
If some CPUs are offline, then make sure they are not included in the
fullMask even if norespect is given to KMP_AFFINITY.
Differential Revision: https://reviews.llvm.org/D112274
Remove restriction forcing users to specify the KMP_HW_SUBSET value in
topology order. This patch sorts the user KMP_HW_SUBSET value before
trying to apply it. For example: 1s,4c,2t is equivalent to 2t,1s,4c
Differential Revision: https://reviews.llvm.org/D112027
The RAII class used for debugging RTL entry used a shared variable to
keep track of the current depth. This used a global initializer, which
isn't supported on AMDGPU. This patch removes the initializer and
instead sets it to zero when the state is initialized in the runtime.
Reviewed By: jdoerfert, JonChesterfield
Differential Revision: https://reviews.llvm.org/D113963
Extension of D112504. Lower amdgpu printf to `__llvm_omp_vprintf`
which takes the same const char*, void* arguments as cuda vprintf and also
passes the size of the void* alloca which will be needed by a non-stub
implementation of `__llvm_omp_vprintf` for amdgpu.
This removes the amdgpu link error on any printf in a target region in favour
of silently compiling code that doesn't print anything to stdout.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112680
The existing CGOpenMPRuntimeAMDGCN and CGOpenMPRuntimeNVPTX classes are
just code bloat. By removing them, the codebase gets a bit cleaner.
Reviewed By: jdoerfert, JonChesterfield, tianshilei1992
Differential Revision: https://reviews.llvm.org/D113421
Have standalone builds define uppercase_CMAKE_BUILD_TYPE and use it.
llvm/CMakeLists.txt defines uppercase_CMAKE_BUILD_TYPE for regular LLVM
builds with OpenMP enabled.
Differential Revision: https://reviews.llvm.org/D112951
The existing CGOpenMPRuntimeAMDGCN and CGOpenMPRuntimeNVPTX classes are
just code bloat. By removing them, the codebase gets a bit cleaner.
Reviewed By: jdoerfert, JonChesterfield, tianshilei1992
Differential Revision: https://reviews.llvm.org/D113421
Extension of D112504. Lower amdgpu printf to `__llvm_omp_vprintf`
which takes the same const char*, void* arguments as cuda vprintf and also
passes the size of the void* alloca which will be needed by a non-stub
implementation of `__llvm_omp_vprintf` for amdgpu.
This removes the amdgpu link error on any printf in a target region in favour
of silently compiling code that doesn't print anything to stdout.
The exact set of changes to check-openmp probably needs revision before commit
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112680
[NFC] As part of using inclusive language within the llvm project,
this patch replaces master with main when referring to `.chm` files.
Reviewed By: teemperor
Differential Revision: https://reviews.llvm.org/D113299
It is better to set all barrier patterns to use "dist" when at least
one environment variable specifies "dist". Otherwise if only one
environment is set to "dist" and others left blank inadvertently,
it would result in mixing dist barrier with default hyper barrier
pattern.
Differential Revision: https://reviews.llvm.org/D112597
Minimize the `impl` interface and clean up some uses of mapping
functions.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D112154
Before we had aligned barriers the `__kmpc_barrier_simple_spmd` was
OK to be used in the custom state machine. Now that SPMD barriers are
assumed to be aligned we need to use a "generic" barrier in places
that are not aligned.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112893
When we pick state 0 to initialize state but thread N is going to be the
"main thread", in generic mode, we would require extra synchronization.
Instead, we should pick the main thread to initialize state in generic
mode and any thread in SPMD mode.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112874
This is a new draft of D28234. I previously did the unorthodox thing of
pushing to it when I wasn't the original author, but since this version
- Uses `GNUInstallDirs`, rather than mimics it, as the original author
was hesitant to do but others requested.
- Is much broader, effecting many more projects than LLVM itself.
I figured it was time to make a new revision.
I am using this patch (and many back-ports) as the basis of
https://github.com/NixOS/nixpkgs/pull/111487 for my distro (NixOS). It
looked like people were generally on board in D28234, but I make note of
this here in case extra motivation is useful.
---
As pointed out in the original issue, a central tension is that LLVM
already has some partial support for these sorts of things. For example
`LLVM_LIBDIR_SUFFIX`, or `COMPILER_RT_INSTALL_PATH`. Because it's not
quite clear yet what to do about those, we are holding off on changing
libdirs and `compiler-rt`. for this initial PR.
---
On the advice of @lebedev.ri, I am splitting this up a bit per
subproject, starting with LLVM. To allow it to be more easily reviewed. This and the subsequent patch must be landed together, as this will not build alone. But the rest can be landed on their own.
Reviewed By: compnerd
Differential Revision: https://reviews.llvm.org/D100810
The synchronization at the end of parallel region cannot make sure all threads
exit the scope. As a result, the assertions right after it might be hit, and
further the `state::assumeInitialState(IsSPMD)` in `__kmpc_target_deinit` may
not hold as well. We either add a synchronization right after the parallel region,
or remove the assertions and assuptions. Here we choose the first one as those
assertions and assumptions can help optimizations.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112861
Summary:
A previous patch changed the check and mistakenly only did `!expr` when
this is a macro expansion and could only apply to the left side of an
expression.
This patch changes the `assert_assume` function used for internal
assumptions in the device runtime to use a more standard formatting for
the assumption message.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112842
Add documentation for the debugging features in the OpenMP device
runtime library.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112010
A common problem is the device running out of global heap memory and
crashing due to a nullptr dereference when using the data sharing stack.
This explicitly checks that a nullptr was not returned by malloc when
debugging field 1 is enabled.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112005
This patch adds support for using function tracing features to track the
executino of runtime functions in the device runtime library. This is
enabled by first compiling the new runtime with
`-fopenmp-target-debug=3` and running with
`LIBOMPTARGET_DEVICE_RTL_DEBUG=3`. The output only tracks team 0 and
thread 0 so there isn't much output when using a generic region.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112002
Passes same tests as the current deviceRTL. Includes cmake change from D111987.
CI is showing a different set of pass/fails to local, committing this
without the tests enabled by default while debugging that difference.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112227
Passes same tests as the current deviceRTL. Includes cmake change from D111987.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D112227
We do not generate _serialized_parallel calls in device mode, no
need for an external API.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D112145
Exiting a data environment will reset all values, it is wrong to adjust
them afterwards.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112144
We will later use the fact that a barrier is aligned to reason about
thread divergence. For now we introduce the assumption and some more
documentation.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112153
The OpenMP thread ID is not the hardware thread ID if we have nesting.
We need to ask the runtime properly to ensure correct results.
Note that the loop interface is going to change soon so we do not adjust
it now but simply ignore the extra argument.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D111950
The team size could/should be an ICV but since we know it is either 1 or
a value we can leave it in the team state for now. However, we still
need to determine if the current level is nested before we use it.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D111949
The first thread state in the new GPU runtime doesn't have a previous
one and we should not dereference the nullptr placeholder.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D111946
According to dlsym description, the value of symbol could be NULL,
and there is no error in this case. Thus dlerror will also return NULL in
this case. We need to check the value returned by dlerror before printing it.
Differential Revision: https://reviews.llvm.org/D112174
Essentially moves the foreach over sm integers into a macro and instantiates it for nvptx.
NFC in that the macro is not presently instantiated for amdgpu as the corresponding code doesn't compile yet.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D111987
Declarations of 5.1 atomic entries were added under
"#if KMP_ARCH_X86 || KMP_ARCH_X86_64" in kmp_atomic.h,
but definitions of the functions missed architecture guard in kmp_atomic.cpp.
As a result mangled symbols were available on non-x86 architecture.
The patch eliminates these unexpected symbols from the library.
Differential Revision: https://reviews.llvm.org/D112261
__ompt_get_task_info_internal function is adapted to support thread_num
determination during the execution of multiple nested serialized
parallel regions enclosed by a regular parallel region.
Consider the following program that contains parallel region R1 executed
by two threads. Let the worker thread T of region R1 executes serialized
parallel regions R2 that encloses another serialized parallel region R3.
Note that the thread T is the master thread of both R2 and R3 regions.
Assume that __ompt_get_task_info_internal function is called with the
argument "ancestor_level == 1" during the execution of region R3.
The function should determine the "thread_num" of the thread T inside
the team of region R2, whose implicit task is at level 1 inside the
hierarchy of active tasks. Since the thread T is the master thread of
region R2, one should expected that "thread_num" takes a value 0.
After the while loop finishes, the following stands: "lwt != NULL",
"prev_lwt == NULL", "prev_team" represents the team information about
the innermost serialized parallel region R3. This results in executing
the assignment "thread_num = prev_team->t.t_master_tid". Note that
"prev_team->t.t_master_tid" was initialized at the moment of
R2’s creation and represents the "thread_num" of the thread T inside
the region R1 which encloses R2. Since the thread T is the worker thread
of the region R1, "the thread_num" takes value 1, which is a contradiction.
This patch proposes to use "lwt" instead of "prev_lwt" when determining
the "thread_num". If "lwt" exists, the task at the requested level belongs
to the serialized parallel region. Since the serialized parallel region
is executed by one thread only, the "thread_num" takes value 0.
Similarly, assume that __ompt_get_task_info_internal function is called
with the argument "ancestor_level == 2" during the execution of region R3.
The function should determine the "thread_num" of the thread T inside the
team of region R1. Since the thread is the worker inside the region R1,
one should expected that "thread_num" takes value 1. After the loop finishes,
the following stands: "lwt == NULL", "prev_lwt != NULL", "prev_team" represents
the team information about the innermost serialized parallel region R3.
This leads to execution of the assignment "thread_num = 0", which causes
a contradiction.
Ignoring the "prev_lwt" leads to executing the assignment
"thread_num = prev_team->t.t_master_tid" instead. From the previous explanation,
it is obvious that "thread_num" takes value 1.
Note that the "prev_lwt" variable is marked as unnecessary and thus removed.
This patch introduces the test case which represents the OpenMP program
described earlier in the summary.
Differential Revision: https://reviews.llvm.org/D110699
__kmp_fork_call sets the enter_frame of the active task (th_curren_task)
before new parallel region begins. After the region is finished, the
enter_frame is cleared.
The old implementation of __kmpc_fork_call didn’t clear the enter_frame of
active task.
Also, the way of initializing the enter_frame of the active task was wrong.
Consider the following two OpenMP programs.
The first program: Let R1 be the serialized parallel region that encloses
another serialized parallel region R2. Assume that thread that executes R2 is
going to create a new serialized parallel region R3 by executing
__kmpc_fork_call. This thread is responsible to set enter_frame of R2's
implicit task. Note that the information about R2's implicit task is present
inside master_th->th.th_current_task at this moment, while lwt represents the
information about R1's implicit task. The old implementation uses lwt and
resets enter_frame of R1's implicit task instead of R2's implicit task. The
new implementation uses master_th->th.th_current_task instead.
The second program: Consider the OpenMP program that contains parallel region
R1 which encloses an explicit task T. Assume that thread should create another
parallel region R2 during the execution of the T. The __kmpc_fork_call is
responsible to create R2 and set enter frame of T whose information is present
inside the master_th->th.th_current_task.
Old implementation tries to set the frame of
parent_team->t.t_implicit_task_taskdata[tid] which corresponds to the implicit
task of the R1, instead of T.
Differential Revision: https://reviews.llvm.org/D112419
As discussed in D108488, testing for invariants of omp_get_wtime would be more
reliable than testing for duration of sleep, as return from sleep might be
delayed due to system load.
Alternatively/in addition, we could compare the time measured by omp_get_wtime
to time measured with C++11 chrono (for portability?).
Differential Revision: https://reviews.llvm.org/D112458
The CHECK: line in the test had no effect, because the test does not
pipe to FileCheck. Since the test only checks for a single value,
encode the result in the return value of the test.
Where possible change to declare the variable before the loop.
Where not possible, specifically request -std=c99 (could be limited to
specific compilers like icc).
Implemented by patching python config instead of modifying all
the tests so that -generic and XFAIL work as usual. Expectation is for
this to be reverted once the old runtime is deleted.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D112225
KMP_API_NAME_GOMP_PARALLEL_SECTIONS function was missing the task frame support.
This patch introduced a fix responsible to set properly the exit_frame of
the innermost implicit task that corresponds to the parallel section construct,
as well as the enter_frame of the task that encloses the mentioned implicit task.
This patch also introduced a simple test case sections_serialized.c that contains
serialized parallel section construct and validates whether the mentioned
task frames are set correctly.
Differential Revision: https://reviews.llvm.org/D112205
Step towards building the DeviceRTL for amdgpu.
Mostly replaces cuda-specific toolchain finding logic with the
generic logic currently found in the amdgpu deviceRTL cmake. Also
deletes dead code and changes the default to build on systems
without cuda installed, as the library doesn't use cuda and the
amdgpu-only systems generally won't have cuda installed.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D111983
The plugin currently uses a macro to check if this is a debug built
before assigning the debug kind variable to the device environment
struct. This is being deprecated because the new device runtime does not
maintain separate debug builds and should always be availible.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D112083
This patch allows to simplify compiler implementation on "taskwait nowait"
construct. The "taskwait nowait" is semantically equivalent to the empty task.
Instead of creating an empty routine as a task entry, compiler can just send
NULL pointer to the runtime. Then the runtime will make all the work with
dependences and return because of the absent task routine.
Differential Revision: https://reviews.llvm.org/D112015
__ompt_get_task_info_internal is now able to determine the right value of the
“thread_num” argument during the execution of an explicit task.
During the execution of a while loop that iterates over the ancestor tasks
hierarchy, the “prev_team” variable was always set to “team” variable at the
beginning of each loop iteration.
Assume that the program contains a parallel region which encloses an explicit
task executed by the worker thread of the region. Also assume that the tool
inquires the “thread_num” of a worker thread for the implicit task that
corresponds to the region (task at “ancestor_level == 1”) and expects to
receive the value of “thread_num > 0”.
After the loop finishes, both “team” and “prev_team” variables are equal and
point to the team information of the parallel region.
The “thread_num” is set to “prev_team->t.t_master_tid”, that is equal to
“team->t.t_master_tid”. In this case, “team->t.t_master_tid” is 0, since
the master thread of the region is the initial master thread of the program.
This leads to a contradiction.
To prevent this, “prev_team” variable is set to “team” variable only at the
time when the loop that has already encountered the implicit task (“taskdata”
variable contains the information about an implicit task) continues iterating
over the implicit task’s ancestors, if any.
After the mentioned loop finishes, the “prev_team” variable might be equal to
NULL. This means that the task at requested “ancestor_level” belongs to the
innermost parallel region, so the “thread_num” will be determined by calling
the “__kmp_get_tid”.
To prove that this patch works, the test case “explicit_task_thread_num.c” is
provided.
It contains the example of the program explained earlier in the summary.
Differential Revision: https://reviews.llvm.org/D110473
Older intel compilers miss the privatization of nested loop variables for
doacross loops. Declaring the variable in the loop makes the test more
robust.
D110279 introduced a bug to the device runtime. In `__kmpc_parallel_51`, we detect
whether we are already in parallel region by `__kmpc_parallel_level() > __kmpc_is_spmd_exec_mode()`.
It is based on the assumption that:
- In SPMD mode, parallel level is initialized to 1.
- In generic mode, parallel level is initialized to 0.
- `__kmpc_is_spmd_exec_mode` returns `1` for SPMD mode, 0 otherwise.
Because the return value type of `__kmpc_is_spmd_exec_mode` is `int8_t`, there
was an implicit cast from `bool` to `int8_t`. We can make sure it is either 0 or
1 since C++14. In D110279, the return value is the result of an `and` operation,
which is 2 in SPMD mode. This breaks the assumption in `__kmpc_parallel_51`.
Reviewed By: carlo.bertolli, dpalermo
Differential Revision: https://reviews.llvm.org/D111905
Detect, through CPUID.1A, and show user different core types through
KMP_AFFINITY=verbose mechanism. Offer future runtime optimizations
__kmp_is_hybrid_cpu() to know whether running on a hybrid system or not.
Differential Revision: https://reviews.llvm.org/D110435
This patch implements teams affinity on the host.
The default is spread. A user can specify either spread, close, or
primary using KMP_TEAMS_PROC_BIND environment variable. Unlike
OMP_PROC_BIND, KMP_TEAMS_PROC_BIND is only a single value and is not a
list of values. The values follow the same semantics under the OpenMP
specification for parallel regions except T is the number of teams in
a league instead of the number of threads in a parallel region.
Differential Revision: https://reviews.llvm.org/D109921
Added functions those implement "atomic compare".
Though clang does not use library interfaces to implement OpenMP atomics,
the functions added for consistency.
Also added missed functions for 80-bit floating min/max atomics.
Differential Revision: https://reviews.llvm.org/D110109
Replaced storing of ittnotify domain array index into
location info structure (which is now read-only) with storing of
(location info address + ittnotify domain + team size) into hash map.
Replaced __kmp_itt_barrier_domains and __kmp_itt_imbalance_domains arrays with
__kmp_itt_barrier_domains hash map; __kmp_itt_region_domains and
__kmp_itt_region_team_size arrays with __kmp_itt_region_domains hash map.
Basic functionality did not change (at least tried to not change).
The patch fixes https://bugs.llvm.org/show_bug.cgi?id=48644.
Differential Revision: https://reviews.llvm.org/D111580
This patch adds support for the
`__kmpc_get_hardware_num_threads_in_block` function that returns the
number of threads. This was missing in the new runtime and was used by
the AMDGPU plugin which prevented it from using the new runtime. This
patchs also unified the interface for getting the thread numbers in the
frontend.
Originally authored by jdoerfert.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D111475
Until we hit the first barrier we should not call `mapping::isSPMDMode`
with all threads. Instead, we now have (and use during initialization) a
`mapping::isMainThreadInGenericMode` overload that takes the known
SPMD-mode state and one that queries it.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D111381
This patch adds an external interface to access the dynamic shared
memory buffer in the device runtime. The function introduced is
``llvm_omp_get_dynamic_shared``. This includes a host-side
definition that only returns a null pointer so that it can be used when
host-fallback is enabled without crashing. Support for dynamic shared
memory was also ported to the old device runtime.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D110957
For NVPTX, `printf` can be used just with a function declaration. For AMDGCN, an
function definition is added, but it simply returns.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D109728
We need to synchronize the threads *before* we destroy the RAII objects
that hold the old values and not after to avoid threads executing the
parallel region but seeing an inconsistent state.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D111369
Follow on to D110006, related to D110957
Where implementations have diverged this resolves to match the new DeviceRTL
- replaces definitions of this struct in deviceRTL and plugins with include
- changes the dynamic_shared_size field from D110006 to 32 bits
- handles stdint being unavailable in DeviceRTL
- adds a zero initializer for the field to amdgpu
- moves the extern declaration for deviceRTL to target_interface
(omptarget.h is more natural, but doesn't work due to include order
with debug.h)
- Renames the fields everywhere to match the LLVM format used in DeviceRTL
- Makes debug_level uint32_t everywhere (previously sometimes int32_t)
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D111069
The hand-rolled linking logic in elf_common does not account for
the possibility of using LLVM dylib rather than a dozen static
libraries. Since it does not seem to be easily convertible
to add_llvm_library, just hand-roll support for LLVM_LINK_LLVM_DYLIB.
This is necessary to support stand-alone builds against installed LLVM.
Differential Revision: https://reviews.llvm.org/D111038
Alexey Bataev