This patch adds:
* New omp_sched_monotonic flag to omp_sched_t which is handled within the runtime
* Parsing of monotonic/nonmonotonic in OMP_SCHEDULE
* Tests for the monotonic flag and envirable parsing
* Logic to force monotonic when hierarchical scheduling is used
Differential Revision: https://reviews.llvm.org/D60979
llvm-svn: 359601
The distribute clause needs an explicit push of a timer. The teams
clause needs a timer added and also, similarly to parallel, exchanged
with the serial timer when encountered so that serial regions are
counted properly.
Differential Revision: https://reviews.llvm.org/D59801
llvm-svn: 357621
This patch cleans up the yielding code and makes it optional. An
environment variable, KMP_USE_YIELD, was added. Yielding is still
on by default (KMP_USE_YIELD=1), but can be turned off completely
(KMP_USE_YIELD=0), or turned on only when oversubscription is detected
(KMP_USE_YIELD=2). Note that oversubscription cannot always be detected
by the runtime (for example, when the runtime is initialized and the
process forks, oversubscription cannot be detected currently over
multiple instances of the runtime).
Because yielding can be controlled by user now, the library mode
settings (from KMP_LIBRARY) for throughput and turnaround have been
adjusted by altering blocktime, unless that was also explicitly set.
In the original code, there were a number of places where a double yield
might have been done under oversubscription. This version checks
oversubscription and if that's not going to yield, then it does
the spin check.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D58148
llvm-svn: 355120
The three switch fallthrough generate a warning with -Wimplicit-fallthrough.
Two are documented as fallthrough, one is not, but I think the intention is to also fallthrough in kmp_tasking.cpp.
Not sure whether kmp.h is the best place to define the macro.
Reviewers: jlpeyton, AndreyChurbanov, Hahnfeld
Reviewed By: jlpeyton
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D56397
llvm-svn: 353052
to reflect the new license. These used slightly different spellings that
defeated my regular expressions.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351648
Add omp_pause_resource and omp_pause_resource_all API and enum, plus stub for
internal implementation. Implemented callable helper function to do local pause,
and added basic functionality for hard and soft pause.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D55078
llvm-svn: 351372
This change includes miscellaneous improvements as follows:
1) Added ompt_get_proc_id() implementation for Windows
2) Added parser and print tool for omp-tool-var, just in case it needs
to be printed (OMP_DISPLAY_ENV)
3) omp_control_tool is exported on Windows
Patch by Hansang Bae
Differential Revision: https://reviews.llvm.org/D50538
llvm-svn: 343137
This patch cleans up unused functions, variables, sign compare issues, and
addresses some -Warning flags which are now enabled including -Wcast-qual.
Not all the warning flags in LibompHandleFlags.cmake are enabled, but some
are with this patch.
Some __kmp_gtid_from_* macros in kmp.h are switched to static inline functions
which allows us to remove the awkward definition of KMP_DEBUG_ASSERT() and
KMP_ASSERT() macros which used the comma operator. This had to be done for the
innumerable -Wunused-value warnings related to KMP_DEBUG_ASSERT()
Differential Revision: https://reviews.llvm.org/D49105
llvm-svn: 339393
1) Remove unnecessary data from list node structure
2) Remove timerPair in favor of pushing/popping explicitTimers.
This way, nested timers will work properly.
3) Fix #pragma omp critical timers
4) Add histogram capability
5) Add KMP_STATS_FILE formatting capability
6) Have time partitioned into serial & parallel by introducing
partitionedTimers::exchange(). This also counts the number of serial regions
in the executable.
7) Fix up the timers around OMP loops so that scheduling overhead and work are
both counted correctly.
8) Fix up the iterations statistics so they count the number of iterations the
thread receives at each loop scheduling event
9) Change timers so there is only one RDTSC read per event change
10) Fix up the outdated comments for the timers
Differential Revision: https://reviews.llvm.org/D49699
llvm-svn: 338276
This patch introduces the logic implementing hierarchical scheduling.
First and foremost, hierarchical scheduling is off by default
To enable, use -DLIBOMP_USE_HIER_SCHED=On during CMake's configure stage.
This work is based off if the IWOMP paper:
"Workstealing and Nested Parallelism in SMP Systems"
Hierarchical scheduling is the layering of OpenMP schedules for different layers
of the memory hierarchy. One can have multiple layers between the threads and
the global iterations space. The threads will go up the hierarchy to grab
iterations, using possibly a different schedule & chunk for each layer.
[ Global iteration space (0-999) ]
(use static)
[ L1 | L1 | L1 | L1 ]
(use dynamic,1)
[ T0 T1 | T2 T3 | T4 T5 | T6 T7 ]
In the example shown above, there are 8 threads and 4 L1 caches begin targeted.
If the topology indicates that there are two threads per core, then two
consecutive threads will share the data of one L1 cache unit. This example
would have the iteration space (0-999) split statically across the four L1
caches (so the first L1 would get (0-249), the second would get (250-499), etc).
Then the threads will use a dynamic,1 schedule to grab iterations from the L1
cache units. There are currently four supported layers: L1, L2, L3, NUMA
OMP_SCHEDULE can now read a hierarchical schedule with this syntax:
OMP_SCHEDULE='EXPERIMENTAL LAYER,SCHED[,CHUNK][:LAYER,SCHED[,CHUNK]...]:SCHED,CHUNK
And OMP_SCHEDULE can still read the normal SCHED,CHUNK syntax from before
I've kept most of the hierarchical scheduling logic inside kmp_dispatch_hier.h
to try to keep it separate from the rest of the code.
Differential Revision: https://reviews.llvm.org/D47962
llvm-svn: 336571
This patch reorganizes the loop scheduling code in order to allow hierarchical
scheduling to use it more effectively. In particular, the goal of this patch
is to separate the algorithmic parts of the scheduling from the thread
logistics code.
Moves declarations & structures to kmp_dispatch.h for easier access in
other files. Extracts the algorithmic part of __kmp_dispatch_init() and
__kmp_dispatch_next() into __kmp_dispatch_init_algorithm() and
__kmp_dispatch_next_algorithm(). The thread bookkeeping logic is still kept in
__kmp_dispatch_init() and __kmp_dispatch_next(). This is done because the
hierarchical scheduler needs to access the scheduling logic without the
bookkeeping logic. To prepare for new pointer in dispatch_private_info_t, a
new flags variable is created which stores the ordered and nomerge flags instead
of them being in two separate variables. This will keep the
dispatch_private_info_t structure the same size.
Differential Revision: https://reviews.llvm.org/D47961
llvm-svn: 336568
Allocated memory is typically not 'const' if it needs to be freed.
This patch removes around 50 wrong const attributes, modifies the
corresponding functions and finally gets rid of some const_casts.
These have especially been strange for __kmp_str_fname_free() that
added a 'const' to call __kmp_str_free() which removed it again.
Two minor cleanups that I performed in this process:
* __kmp_tool_libraries now lives in kmp_settings.cpp as it is
used nowhere else.
* __kmp_msg_empty was removed as it was never used and Clang
now complained that it was assigned a string literal that
is 'const char *'.
Differential Revision: https://reviews.llvm.org/D39755
llvm-svn: 317797
The TR6 document is expected to be publically released around November 15.
This patch does not implement OMPT for libomptarget.
Patch by Simon Convent and Joachim Protze
Differential Revision: https://reviews.llvm.org/D39182
llvm-svn: 317339
The code is tested to work with latest clang, GNU and Intel compiler. The implementation
is optimized for low overhead when no tool is attached shifting the cost to execution with
tool attached.
This patch does not implement OMPT for libomptarget.
Patch by Simon Convent and Joachim Protze
Differential Revision: https://reviews.llvm.org/D38185
llvm-svn: 317085
Removes semicolons after if {} blocks, function definitions, etc.
I was able to apply the large OMPT patch cleanly on top of this one
with no conflicts.
llvm-svn: 314340
Minor code cleanup of Klocwork issues. Fatal messages are given no return
attribute. Define and use KMP_NORETURN to work for multiple C++ versions.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D37275
llvm-svn: 312538
Changes are: got all atomics to accept volatile pointers that allowed
to simplify many type conversions. Windows specific code fixed correspondingly.
Differential Revision: https://reviews.llvm.org/D35417
llvm-svn: 308164
While importing libomp into the FreeBSD base system we encountered
Clang warnings that "'register' storage class specifier is deprecated
and incompatible with C++1z [-Wdeprecated-register]".
Differential Revision: https://reviews.llvm.org/D35124
llvm-svn: 307441
Changes are: replaced C-style casts with cons_cast and reinterpret_cast;
type of several counters changed to signed; type of parameters of 32-bit and
64-bit AND and OR intrinsics changes to unsigned; changed files formatted
using clang-format version 3.8.1.
Differential Revision: https://reviews.llvm.org/D34759
llvm-svn: 307020
This patch contains the clang-format and cleanup of the entire code base. Some
of clang-formats changes made the code look worse in places. A best effort was
made to resolve the bulk of these problems, but many remain. Most of the
problems were mangling line-breaks and tabbing of comments.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D32659
llvm-svn: 302929
Add counter to count number of static_steal for loops
Add counter for number of chunks executed per static_steal for loop
Add counter for number of chunks stolen per static_steal for loop
llvm-svn: 295461
Put the duplicated i_maxmin into traits_t by adding new members max_value and
min_value. Put ___kmp_size_type into traits_t by adding member type_size.
Differential Revision: https://reviews.llvm.org/D28847
llvm-svn: 293316
When linking with libhwloc, the ORDERED EPCC test slows down on big
machines (> 48 cores). Performance analysis showed that a cache thrash
was occurring and this padding helps alleviate the problem.
Also, inside the main spin-wait loop in kmp_wait_release.h, we can eliminate
the references to the global shared variables by instead creating a local
variable, oversubscribed and instead checking that.
Differential Revision: http://reviews.llvm.org/D22093
llvm-svn: 274894
Replaced readings of nproc from team structure with ones from
thread structure to improve performance.
Patch by Andrey Churbanov.
Differential Revision: http://reviews.llvm.org/D21559
llvm-svn: 273293
OpenMP 4.1 is now OpenMP 4.5. Any mention of 41 or 4.1 is replaced with
45 or 4.5. Also, if the CMake option LIBOMP_OMP_VERSION is 41, CMake warns that
41 is deprecated and to use 45 instead.
llvm-svn: 272687
Fix for bugzilla https://llvm.org/bugs/show_bug.cgi?id=26602. Removed functions
body consisted of the only KMP_ASSERT(0) statement. Thus possible runtime crash
converted to compile-time error, which looks preferable (faster possible error
detection).
TODO: consider C++11 static assert as an alternative, that could
make the diagnostics better.
Patch by Andrey Churbanov
Differential Revision: http://reviews.llvm.org/D21304
llvm-svn: 272590
This patch replaces use of compiler builtin atomics with
C++11 atomics for ticket locks implementation. Ticket locks
are used in critical places of the runtime, e.g. in the tasking
mechanism.
The main reason this change was introduced is the problem
with work stealing function on ARM architecture which suffered
from nasty race condition. It turned out that the root cause of
the problem lies in the way ticket locks are implemented. Changing
compiler builtins into C++11 atomics solves the problem.
Two assertions were added into kmp_tasking.c which are useful
for detecting early symptoms of something wrong going on with
work stealing, which were among the possible outcomes of the
race condition.
Differential Revision: http://reviews.llvm.org/D19878
llvm-svn: 271324
The problem is the lack of dispatch buffers when thousands of loops with nowait,
about 10 iterations each, are executed by hundreds of threads. We only have
built-in 7 dispatch buffers, but there is a need in dozens or hundreds of
buffers.
The problem can be fixed by setting KMP_MAX_DISP_BUF to bigger value. In order
to give users same possibility I changed build-time control into run-time one,
adding API just in case.
This change adds an environment variable KMP_DISP_NUM_BUFFERS and a new API
function kmp_set_disp_num_buffers(int num_buffers).
The KMP_DISP_NUM_BUFFERS envirable works only before serial initialization,
because during the serial initialization we already allocate buffers for the hot
team, so it is too late to change the number of buffers later (or we need to
reallocate buffers for all teams which sounds too complicated). The
kmp_set_defaults() routine does not work for this envirable, because it calls
serial initialization before reading the parameter string. So a new routine,
kmp_set_disp_num_buffers(), is created so that it can set our internal global
variable before the library initialization. If both the envirable and API used
the envirable wins.
Differential Revision: http://reviews.llvm.org/D20697
llvm-svn: 271318
This change removes the current timers with ones that partition time properly.
The current timers are nested, so that if a new timer, B, starts when the
current timer, A, is already timing, A's time will include B's. To eliminate
this problem, the partitioned timers are designed to stop the current timer (A),
let the new timer run (B), and when the new timer is finished, restart the
previously running timer (A). With this partitioning of time, a threads' timers
all sum up to the OMP_worker_thread_life time and can now easily show the
percentage of time a thread is spending in different parts of the runtime or
user code.
There is also a new state variable associated with each thread which tells where
it is executing a task. This corresponds with the timers: OMP_task_*, e.g., if
time is spent in OMP_task_taskwait, then that thread executed tasks inside a
#pragma omp taskwait construct.
The changes are mostly changing the MACROs to use the new PARITIONED_* macros,
the new partitionedTimers class and its methods, and new state logic.
Differential Revision: http://reviews.llvm.org/D19229
llvm-svn: 268640
The trip count calculation was incorrect for loops with large bounds. For example,
for(int i=-2,000,000,000; i < 2,000,000,000; i+=50000000), the trip count
calculation had overflow (trying to calculate 2,000,000,000 + 2,000,000,000 with
signed integers) and wasn't giving the right value. This patch fixes this error
in the runtime by using unsigned integers instead. There is still a bug in the
clang compiler component because it warns that there is overflow in the
test case file when there isn't. This error isn't there for the Intel Compiler.
So for now, the test case is designated as XFAIL.
Differential Revision: http://reviews.llvm.org/D19078
llvm-svn: 266677
From the standard: A doacross loop nest is a loop nest that has cross-iteration
dependence. An iteration is dependent on one or more lexicographically earlier
iterations. The ordered clause parameter on a loop directive identifies the
loop(s) associated with the doacross loop nest.
The init/fini routines allocate/free doacross buffer(s) for each loop for each
thread. The wait routine waits for a flag designated by the dependence vector.
The post routine sets the flag designated by current iteration vector. We use
a similar technique of shared buffer indices that covers up to 7 nowait loops
executed simultaneously by different threads (number 7 has no real meaning,
just heuristic value). Also, the size of structures are kept intact via
reducing dummy arrays.
This needs to be put into the OpenMP runtime library in order for the compiler
team to develop the compiler side of the implementation.
Differential Revision: http://reviews.llvm.org/D17399
llvm-svn: 262532
The monotonic/non-monotonic flags are sent to the runtime via the sched_type by
setting the 30th (non-monotonic) or 29th (monotonic) bit in the sched_type.
Macros are added to probe if monotonic or non-monotonic is specified
(SCHEDULE_HAS_[NON]MONOTONIC & SCHEDULE_HAS_NO_MODIFIERS)
and also to to get the base sched_type (SCHEDULE_WITHOUT_MODIFIERS)
Currently, nothing is done with the modifiers.
Also, this patch adds some comments on the use of the enumerations in at least
one place where it is subtle.
Differential Revision: http://reviews.llvm.org/D17406
llvm-svn: 261906
Fix ittnotify loop metadata reporting for schedule(runtime) and
chunked schedule set via OMP_SCHEDULE. The bug was that chunk=1
reported always.
llvm-svn: 252952
Setting dynamic schedule with chunk size 0 via omp_set_schedule(dynamic,0)
and then using "schedule (runtime)" causes infinite loop because for the
chunked dynamic schedule we didn't correct zero chunk to the default (1).
llvm-svn: 252338
Prior to this change, OMPT had a status flag ompt_status, which could take
several values. This was due to an earlier OMPT design that had several levels
of enablement (ready, disabled, tracking state, tracking callbacks). The
current OMPT design has OMPT support either on or off.
This revision replaces ompt_status with a boolean flag ompt_enabled, which
simplifies the runtime logic for OMPT.
Patch by John Mellor-Crummey
Differential Revision: http://reviews.llvm.org/D12999
llvm-svn: 248189
This removes some statistics counters and timers which were not used,
adds new counters and timers for some language features that were not
monitored previously and separates the counters and timers into those
which are of interest for investigating user code and those which are
only of interest to the developer of the runtime itself.
The runtime developer statistics are now ony collected if the
additional #define KMP_DEVELOPER_STATS is set.
Additional user statistics which are now collected include:
* Count of nested parallelism (omp parallel inside a parallel region)
* Count of omp distribute occurrences
* Count of omp teams occurrences
* Counts of task related statistics (taskyield, task execution, task
cancellation, task steal)
* Values passed to omp_set_numtheads
* Time spent in omp single and omp master
None of this affects code compiled without stats gathering enabled,
which is the normal library build mode.
This also fixes the CMake build by linking to the standard c++ library
when building the stats library as it is a requirement. The normal library
does not have this requirement and its link phase is left alone.
Differential Revision: http://reviews.llvm.org/D11759
llvm-svn: 244677
As an ongoing effort to sanitize the openmp code, these changes remove unused variables
by adding proper macros around both variables and functions.
Patch by Jack Howarth
llvm-svn: 239330
