Restructured dynamic loop dispatcher code.
Fixed use of dispatch buffers for nonmonotonic dynamic (static_steal) schedule:
- eliminated possibility of stealing iterations of the wrong loop when victim
thread changed its buffer to work on another loop;
- fixed race when victim thread changed its buffer to work in nested parallel;
- eliminated "static" property of the schedule, that is now a single thread can
execute whole loop.
Differential Revision: https://reviews.llvm.org/D103648
Two-level distributed barrier is a new experimental barrier designed
for Intel hardware that has better performance in some cases than the
default hyper barrier.
This barrier is designed to handle fine granularity parallelism where
barriers are used frequently with little compute and memory access
between barriers. There is no need to use it for codes with few
barriers and large granularity compute, or memory intensive
applications, as little difference will be seen between this barrier
and the default hyper barrier. This barrier is designed to work
optimally with a fixed number of threads, and has a significant setup
time, so should NOT be used in situations where the number of threads
in a team is varied frequently.
The two-level distributed barrier is off by default -- hyper barrier
is used by default. To use this barrier, you must set all barrier
patterns to use this type, because it will not work with other barrier
patterns. Thus, to turn it on, the following settings are required:
KMP_FORKJOIN_BARRIER_PATTERN=dist,dist
KMP_PLAIN_BARRIER_PATTERN=dist,dist
KMP_REDUCTION_BARRIER_PATTERN=dist,dist
Branching factors (set with KMP_FORKJOIN_BARRIER, KMP_PLAIN_BARRIER,
and KMP_REDUCTION_BARRIER) are ignored by the two-level distributed
barrier.
Differential Revision: https://reviews.llvm.org/D103121
This is the first of seven patches that implements OMPD, a debugging interface to support debugging of OpenMP programs.
It contains support code required in "openmp/runtime" for OMPD implementation.
Reviewed By: @hbae
Differential Revision: https://reviews.llvm.org/D100181
Nesting mode is a new experimental feature in the OpenMP
runtime. It allows a user to set up nesting for an application in a
way that corresponds to the hardware topology levels on the machine an
application is being run on. For example, if a machine has 2 sockets,
each with 12 cores, then use of nesting mode could set up an outer
level of nesting that uses 2 threads per parallel region, and an inner
level of nesting that uses 12 threads per parallel region.
Nesting mode is controlled with the KMP_NESTING_MODE environment
variable as follows:
1) KMP_NESTING_MODE = 0: Nesting mode is off (default); max-active-levels-var
is set to 1 (the default -- nesting is off, nested parallel regions
are serialized).
2) KMP_NESTING_MODE = 1: Nesting mode is on, and a number of threads
will be assigned for each level discovered in the machine topology;
max-active-levels-var is set to the number of levels discovered.
3) KMP_NESTING_MODE = n, n>1: [Note: this option is experimental and may change
or be removed in the future.] Nesting mode is on, and a number of
threads will be assigned for each topology level discovered on the
machine, up to k<=n levels (since there may be fewer than n levels
discovered in the topology), and beyond the kth level, nested parallel
regions will be serialized; NOTE: max-active-levels-var is 1 (the default --
nesting is off, and nested parallel regions are serialized until the
user changes max-active-levels-var.
If the user sets OMP_NUM_THREADS or OMP_MAX_ACTIVE_LEVELS, they will
override KMP_NESTING_MODE settings for the associated environment
variables. The detected topology may be limited by an affinity mask
setting on the initial thread, or if the user sets KMP_HW_SUBSET. See
also: KMP_HOT_TEAMS_MAX_LEVEL for controlling use of hot teams for
nested parallel regions. Note that this feature only sets numbers of
threads used at nesting levels. The user should make use of
OMP_PLACES and OMP_PROC_BIND or KMP_AFFINITY for affinitizing those
threads, if desired.
Differential Revision: https://reviews.llvm.org/D102188
This patch does the following:
1) Introduce kmp_topology_t as the runtime-friendly structure (the
corresponding global variable is __kmp_topology) to determine the
exact machine topology which can vary widely among current and future
architectures. The current design is not easy to expand beyond the assumed
three layer topology: sockets, cores, and threads so a rework capable of
using the existing KMP_AFFINITY mechanisms is required.
This new topology structure has:
* The depth and types of the topology
* Ratio count for each consecutive level (e.g., number of cores per
socket, number of threads per core)
* Absolute count for each level (e.g., 2 sockets, 16 cores, 32 threads)
* Equivalent topology layer map (e.g., Numa domain is equivalent to
socket, L1/L2 cache equivalent to core)
* Whether it is uniform or not
The hardware threads are represented with the kmp_hw_thread_t
structure. This structure contains the ids (e.g., socket 0, core 1,
thread 0) and other information grabbed from the previous Address
structure. The kmp_topology_t structure contains an array of these.
2) Generalize the KMP_HW_SUBSET envirable for the new
kmp_topology_t structure. The algorithm doesn't assume any order with
tiles,numa domains,sockets,cores,threads. Instead it just parses the
envirable, makes sure it is consistent with the detected topology
(including taking into account equivalent layers) and then trims away
the unneeded subset of hardware threads. To enable this, a new
kmp_hw_subset_t structure is introduced which contains a vector of
items (hardware type, number user wants, offset). Any keyword within
__kmp_hw_get_keyword() can be used as a name and can be shortened as
well. e.g.,
KMP_HW_SUBSET=1s,2numa,4tile,2c,3t can be used on the KNL SNC-4 machine.
3) Simplify topology detection functions so they only do the singular
task of detecting the machine's topology. Printing, and all
canonicalizing functionality is now done afterwards. So many lines of
duplicated code are eliminated.
4) Add new ll_caches and numa_domains to OMP_PLACES, and
consequently, KMP_AFFINITY's granularity setting. All the names within
__kmp_hw_get_keyword() are available for use in OMP_PLACES or
KMP_AFFINITY's granularity setting.
5) Simplify and future-proof code where explicit lists of allowed
affinity settings keywords inside if() conditions.
6) Add x86 CPUID leaf 4 cache detection to existing x2apic id method
so equivalent caches could be detected (in particular for the ll_caches
place).
Differential Revision: https://reviews.llvm.org/D100997
-- Added or moved checks to appropriate places.
-- Removed ineffective null check where the pointer is already being
dereferenced around the code.
-- Initialized variables that can be used without definitions.
-- Added call to dlclose/FreeLibrary in OMPT tool activation.
-- Added a new build compiler definition.
Differential Revision: https://reviews.llvm.org/D98584
It is reported that after enabling hidden helper thread, the program
can hit the assertion `new_gtid < __kmp_threads_capacity` sometimes. The root
cause is explained as follows. Let's say the default `__kmp_threads_capacity` is
`N`. If hidden helper thread is enabled, `__kmp_threads_capacity` will be offset
to `N+8` by default. If the number of threads we need exceeds `N+8`, e.g. via
`num_threads` clause, we need to expand `__kmp_threads`. In
`__kmp_expand_threads`, the expansion starts from `__kmp_threads_capacity`, and
repeatedly doubling it until the new capacity meets the requirement. Let's
assume the new requirement is `Y`. If `Y` happens to meet the constraint
`(N+8)*2^X=Y` where `X` is the number of iterations, the new capacity is not
enough because we have 8 slots for hidden helper threads.
Here is an example.
```
#include <vector>
int main(int argc, char *argv[]) {
constexpr const size_t N = 1344;
std::vector<int> data(N);
#pragma omp parallel for
for (unsigned i = 0; i < N; ++i) {
data[i] = i;
}
#pragma omp parallel for num_threads(N)
for (unsigned i = 0; i < N; ++i) {
data[i] += i;
}
return 0;
}
```
My CPU is 20C40T, then `__kmp_threads_capacity` is 160. After offset,
`__kmp_threads_capacity` becomes 168. `1344 = (160+8)*2^3`, then the assertions
hit.
Reviewed By: protze.joachim
Differential Revision: https://reviews.llvm.org/D98838
and __kmpc_end_masked. The "master" construct is deprecated. Changed
proc-bind keyword from "master" to "primary". Use of both master
construct and master as proc-bind keyword is still allowed, but
deprecated.
Remove references to "master" in comments and strings, and replace
with "primary" or "primary thread". Function names and variables were
not touched, nor were references to deprecated master construct. These
can be updated over time. No new code should refer to master.
This patch limits the number of dispatch buffers (used for
loop worksharing construct) to between 1 and 4096.
Differential Revision: https://reviews.llvm.org/D96749
This patch introduces a new environment variable to force monotonic
behavior for users that absolutely need it. This is in anticipation
of 5.0 change that uses non-monotonic behavior for dynamic scheduling
by default. Fixes for that and the actual switch are coming soon.
Differential Revision: https://reviews.llvm.org/D95263
When OMP_PLACES contains an invalid value, the warning informs the user
that the fallback is OMP_PLACES=threads, but the actual internal setting
is OMP_PLACES=cores and is detected as such with KMP_SETTINGS=1.
This patch informs the user that OMP_PLACES=cores is being used instead
of OMP_PLACES=threads.
Differential Revision: https://reviews.llvm.org/D95170
This patch adds the new algorithm for topology discovery using cpuid
leaf 1f. Only the new die level is detected and integrated into the
current affinity mechanisms including KMP_AFFINITY (granularity level
and compact/scatter algorithm), OMP_PLACES=dies, and KMP_HW_SUBSET.
Differential Revision: https://reviews.llvm.org/D95157
This patch sets the def-allocator-var ICV based on the environment variables
provided in OMP_ALLOCATOR. Previously, only allowed value for OMP_ALLOCATOR
was a predefined memory allocator. OpenMP 5.1 specification allows predefined
memory allocator, predefined mem space, or predefined mem space with traits in
OMP_ALLOCATOR. If an allocator can not be created using the provided environment
variables, the def-allocator-var is set to omp_default_mem_alloc.
Differential Revision: https://reviews.llvm.org/D94985
The basic design is to create an outer-most parallel team. It is not a regular team because it is only created when the first hidden helper task is encountered, and is only responsible for the execution of hidden helper tasks. We first use `pthread_create` to create a new thread, let's call it the initial and also the main thread of the hidden helper team. This initial thread then initializes a new root, just like what RTL does in initialization. After that, it directly calls `__kmpc_fork_call`. It is like the initial thread encounters a parallel region. The wrapped function for this team is, for main thread, which is the initial thread that we create via `pthread_create` on Linux, waits on a condition variable. The condition variable can only be signaled when RTL is being destroyed. For other work threads, they just do nothing. The reason that main thread needs to wait there is, in current implementation, once the main thread finishes the wrapped function of this team, it starts to free the team which is not what we want.
Two environment variables, `LIBOMP_NUM_HIDDEN_HELPER_THREADS` and `LIBOMP_USE_HIDDEN_HELPER_TASK`, are also set to configure the number of threads and enable/disable this feature. By default, the number of hidden helper threads is 8.
Here are some open issues to be discussed:
1. The main thread goes to sleeping when the initialization is finished. As Andrey mentioned, we might need it to be awaken from time to time to do some stuffs. What kind of update/check should be put here?
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D77609
The basic design is to create an outer-most parallel team. It is not a regular team because it is only created when the first hidden helper task is encountered, and is only responsible for the execution of hidden helper tasks. We first use `pthread_create` to create a new thread, let's call it the initial and also the main thread of the hidden helper team. This initial thread then initializes a new root, just like what RTL does in initialization. After that, it directly calls `__kmpc_fork_call`. It is like the initial thread encounters a parallel region. The wrapped function for this team is, for main thread, which is the initial thread that we create via `pthread_create` on Linux, waits on a condition variable. The condition variable can only be signaled when RTL is being destroyed. For other work threads, they just do nothing. The reason that main thread needs to wait there is, in current implementation, once the main thread finishes the wrapped function of this team, it starts to free the team which is not what we want.
Two environment variables, `LIBOMP_NUM_HIDDEN_HELPER_THREADS` and `LIBOMP_USE_HIDDEN_HELPER_TASK`, are also set to configure the number of threads and enable/disable this feature. By default, the number of hidden helper threads is 8.
Here are some open issues to be discussed:
1. The main thread goes to sleeping when the initialization is finished. As Andrey mentioned, we might need it to be awaken from time to time to do some stuffs. What kind of update/check should be put here?
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D77609
This patch partially prepares the runtime source code to be built with
-Wconversion, which should trigger warnings if any implicit conversions
can possibly change a value. For builds done with icc or gcc, all such
warnings are handled in this patch. clang gives a much longer list of
warnings, particularly for sign conversions, which the other compilers
don't report. The -Wconversion flag is commented into cmake files, but
I'm not going to turn it on. If someone thinks it is important, and wants
to fix all the clang warnings, they are welcome to.
Types of changes made here involve either improving the consistency of types
used so that no conversion is needed, or else performing careful explicit
conversions, when we're sure a problem won't arise.
Patch is a combination of changes by Terry Wilmarth and Johnny Peyton.
Differential Revision: https://reviews.llvm.org/D92942
This patch introduces a new RTM lock type based on spin lock which is
used for OMP lock with speculative hint on supported architecture.
Differential Revision: https://reviews.llvm.org/D92615
These changes add support for Intel's umonitor/umwait usage in wait
code, for architectures that support those intrinsic functions. Usage of
umonitor/umwait is off by default, but can be turned on by setting the
KMP_USER_LEVEL_MWAIT environment variable.
Differential Revision: https://reviews.llvm.org/D91189
OpenMP 5.1 introduces the new env variable
OMP_TOOL_VERBOSE_INIT=(disabled|stdout|stderr|<filename>) to enable verbose
loading and initialization of OMPT tools.
This env variable helps to understand the cause when loading of a tool fails
(e.g., undefined symbols or dependency not in LD_LIBRARY_PATH)
Output of OMP_TOOL_VERBOSE_INIT is added for OMP_DISPLAY_ENV
Tests for this patch are integrated into the different existing tool loading
tests, making these tests more verbose. An Archer specific verbose test is
integrated into an existing Archer test.
Patch prepared by: Isabel Thärigen
Differential Revision: https://reviews.llvm.org/D91464
Remove all older OMP spec versioning from the runtime and build system.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D64534
llvm-svn: 365963
Current parsing allows trailing string after the permitted value,
MANDATORY|DISABLED|DEFAULT -- e.g., "mandatorynot" is also recognized
as "MANDATORY". Such cases should be recognized as incorrect/unknown
value.
Differential Revision: https://reviews.llvm.org/D62431
llvm-svn: 362125
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
* Replace HBWMALLOC API with more general MEMKIND API, new functions
and variables added.
* Have libmemkind.so loaded when accessible.
* Redirect memspaces to default one except for high bandwidth which
is processed separately.
* Ignore some allocator traits e.g., sync_hint, access, pinned, while
others are processed normally e.g., alignment, pool_size, fallback,
fb_data, partition.
* Add tests for memory management
Patch by Andrey Churbanov
Differential Revision: https://reviews.llvm.org/D59783
llvm-svn: 357929
Nest-var, OMP_NESTED, omp_set_nested()., and omp_get_nested() have been
deprecated in the 5.0 spec. Initial nesting info is now derived from
OMP_MAX_ACTIVE_LEVELS, OMP_NUM_THREADS, and OMP_PROC_BIND.
This patch deprecates the internal ICV that corresponds to nest-var, and
replaces it with the max-active-levels-var ICV to determine nesting. The
change still allows for use of OMP_NESTED (according to 5.0 changes),
omp_get_nested, and omp_set_nested, which have had deprecation messages
added to them. The change allows certain settings of OMP_NUM_THREADS,
OMP_PROC_BIND, and OMP_MAX_ACTIVE_LEVELS to turn on nesting, but
OMP_NESTED=0 will still force nesting to be off.
The runtime now prints informative messages about deprecation of
OMP_NESTED, omp_set_nested(), and omp_get_nested(), when those
environment variables or routines are used. It also prints deprecated
message in output for KMP_SETTINGS and OMP_DISPLAY_ENV for OMP_NESTED.
This patch also fixes OMP_DISPLAY_ENV output for OMP_TARGET_OFFLOAD.
Patch by Terry Wilmarth
Differential Revision: https://reviews.llvm.org/D58408
llvm-svn: 355138
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
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
This patch adds the affinity format functionality introduced in OpenMP 5.0.
This patch adds: Two new environment variables:
OMP_DISPLAY_AFFINITY=TRUE|FALSE
OMP_AFFINITY_FORMAT=<string>
and Four new API:
1) omp_set_affinity_format()
2) omp_get_affinity_format()
3) omp_display_affinity()
4) omp_capture_affinity()
The affinity format functionality has two ICV's associated with it:
affinity-display-var (bool) and affinity-format-var (string).
The affinity-display-var enables/disables the functionality through the
envirable OMP_DISPLAY_AFFINITY. The affinity-format-var is a formatted
string with the special field types beginning with a '%' character
similar to printf
For example, the affinity-format-var could be:
"OMP: host:%H pid:%P OStid:%i num_threads:%N thread_num:%n affinity:{%A}"
The affinity-format-var is displayed by every thread implicitly at the beginning
of a parallel region when any thread's affinity has changed (including a brand
new thread being spawned), or explicitly using the omp_display_affinity() API.
The omp_capture_affinity() function can capture the affinity-format-var in a
char buffer. And omp_set|get_affinity_format() allow the user to set|get the
affinity-format-var explicitly at runtime. omp_capture_affinity() and
omp_get_affinity_format() both return the number of characters needed to hold
the entire string it tried to make (not including NULL character). If not
enough buffer space is available,
both these functions truncate their output.
Differential Revision: https://reviews.llvm.org/D55148
llvm-svn: 349089
* Fix calculation of string length.
* Remove NULL-check of pointer which has been dereferenced.
Patch by Andrey Churbanov
Differential Revision: https://reviews.llvm.org/D54948
llvm-svn: 347801
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
Implemented omp_alloc, omp_free, omp_{set,get}_default_allocator entries,
and OMP_ALLOCATOR environment variable.
Added support for HBW memory on Linux if libmemkind.so library is accessible
(dynamic library only, no support for static libraries).
Only used stable API (hbwmalloc) of the memkind library
though we may consider using experimental API in future.
The ICV def-allocator-var is implemented per implicit task similar to
place-partition-var. In the absence of a requested allocator, the uses the
default allocator.
Predefined allocators (the only ones currently available) are made similar
for C and Fortran, - pointers (long integers) with values 1 to 8.
Patch by Andrey Churbanov
Differential Revision: https://reviews.llvm.org/D51232
llvm-svn: 341687
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
This change fixes build errors when building a runtime with adaptive lock stats
enabled. Most of the errors were due to the recent changes in the runtime, but
it seems that we have not tried to build this debug runtime on Windows for a
long time.
Patch by Hansang Bae
Differential Revision: https://reviews.llvm.org/D49823
llvm-svn: 338277
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
Currently, the affinity API reports garbage for the initial place list and any
thread's place lists when using KMP_AFFINITY=none|compact|scatter.
This patch does two things:
for KMP_AFFINITY=none, Creates a one entry table for the places, this way, the
initial place list is just a single place with all the proc ids in it. We also
set the initial place of any thread to 0 instead of KMP_PLACE_ALL so that the
thread reports that single place (place 0) instead of garbage (-1) when using
the affinity API.
When non-OMP_PROC_BIND affinity is used
(including KMP_AFFINITY=compact|scatter), a thread's place list is populated
correctly. We assume that each thread is assigned to a single place. This is
implemented in two of the affinity API functions
Differential Revision: https://reviews.llvm.org/D45527
llvm-svn: 330283