This patch enables balanced affinity on machines that do not have
hardware threads and have cores clustered into packages. In facts,
balacing algorithm could be generalized for any arrangement with
at least two levels of hierarchy (depth > 1).
Differential Revision: https://reviews.llvm.org/D22365
llvm-svn: 277212
A couple improvements:
1) Add ability to limit fullMask size when KMP_HW_SUBSET limits resources.
2) Make KMP_HW_SUBSET work for affinity_none, and only limit fullMask in this case.
Patch by Andrey Churbanov.
Differential Revision: http://reviews.llvm.org/D21528
llvm-svn: 273278
Change hwloc discovery algorithm to print topology for only accessible
resources, and report uniformity correspondingly, similar to what other topology
discovery algorithms do. Fixes minor inconsistency in total topology reported
and resources used for threads binding in case hwloc used.
Patch by Andrey Churbanov.
Differential Revision: http://reviews.llvm.org/D21389
llvm-svn: 272952
Cleanup: fixed missing memory cleanup in couple of corner cases. Fixes possible
memory leak in some corner cases
Patch by Andrey Churbanov
Differential Revision: http://reviews.llvm.org/D21355
llvm-svn: 272946
Deprecate KMP_PLACE_THREADS and rename it to KMP_HW_SUBSET due to confusion
about its purpose and function among users. KMP_HW_SUBSET is an environment
variable which allows users to easily pick a subset of the hardware topology to
use. e.g., KMP_HW_SUBSET=30c,2t means use 30 cores, 2 threads per core.
Patch by Andrey Churbanov
Differential Revision: http://reviews.llvm.org/D21340
llvm-svn: 272937
Remove static specifier from var fullMask and remove kmp_get_fullMask() routine.
When iterating through procs in a mask, always check if proc is in fullMask
(this check was missing in a few places).
Patch by Brian Bliss.
Differential Revision: http://reviews.llvm.org/D21300
llvm-svn: 272589
These changes remove the hwloc_topology_ignore_type function which doesn't exist
in the hwloc 2.0 API. In the existing code, the topology extracted from hwloc
has the cache levels stripped out and then assumes the final stripped topology
follows the typical three-level topology: packages -> cores -> HW threads.
But the code is doing unclean manipulations to determine at what level those
resources are located and also assumes too much about what hwloc is detecting
(there could be intermediate levels in between socket and core for instance).
This new way of extracting the topology doesn't strip out any hardware objects
that hwloc detects. It does not assume the three level topology, and instead
searches for the relevant three levels within the topology for each bit of
information using hwloc interface functions. i.e., the three level topology
subset that our affinity code is interested in is extracted from the hwloc
topology tree directly.
For example, the new __kmp_hwloc_get_nobjs_under_obj function gives the user the
number of cores under a socket reliably without worrying if there are unexpected
objects between the socket object and core object in the hwloc topology
structure. Also, now that all topology information is kept, there are also
possibilities of using the caches/numa nodes to determine more sophisticated
affinity settings in the future.
There is also some cleanup code added for the destruction of the
__kmp_hwloc_topology object.
Differential Revision: http://reviews.llvm.org/D21195
llvm-svn: 272565
This debug sections's functionality can be replicated using the environment
variable KMP_TOPOLOGY_METHOD with different values and KMP_AFFINITY=verbose
llvm-svn: 267472
This change has the hwloc_bitmap_list_snprintf() function use the entire buffer
to print the mask. There is no need to shorten the buffer length by 7. It only
needs to be shortened by one byte.
llvm-svn: 267470
This new API, int kmp_set_thread_affinity_mask_initial(), is available for use
by other parallel runtime libraries inside a possibly OpenMP-registered thread.
This entry point restores the current thread's affinity mask to the affinity
mask of the application when it first began. If -1 is returned it can be assumed
that either the thread hasn't called affinity initialization or that the thread
isn't registered with the OpenMP library. If 0 is returned then, then the call
was successful. Any return value greater than zero indicates an error occurred
when setting affinity.
Differential Revision: http://reviews.llvm.org/D15867
llvm-svn: 257489
These changes allow libhwloc to be used as the topology discovery/affinity
mechanism for libomp. It is supported on Unices. The code additions:
* Canonicalize KMP_CPU_* interface macros so bitmask operations are
implementation independent and work with both hwloc bitmaps and libomp
bitmaps. So there are new KMP_CPU_ALLOC_* and KMP_CPU_ITERATE() macros and
the like. These are all in kmp.h and appropriately placed.
* Hwloc topology discovery code in kmp_affinity.cpp. This uses the hwloc
interface to create a libomp address2os object which the rest of libomp knows
how to handle already.
* To build, use -DLIBOMP_USE_HWLOC=on and
-DLIBOMP_HWLOC_INSTALL_DIR=/path/to/install/dir [default /usr/local]. If CMake
can't find the library or hwloc.h, then it will tell you and exit.
Differential Revision: http://reviews.llvm.org/D13991
llvm-svn: 254320
These changes include:
1) Machine hierarchy now uses the base_num_threads field to indicate the
maximum number of threads the current hierarchy can handle without a resize.
2) In __kmp_get_hierarchy, we need to get depth after any potential resize
is done.
3) Cleanup of hierarchy resize code to support 1 above.
Differential Revision: http://reviews.llvm.org/D14455
llvm-svn: 252475
Added (optional) sockets to the syntax of the KMP_PLACE_THREADS environment variable.
Some limitations:
* The number of sockets and then optional offset should be specified first (before other parameters).
* The letter designation is mandatory for sockets and then for other parameters.
* If number of cores is specified first, then the number of sockets is defaulted to all sockets on the machine; also, the old syntax is partially supported if sockets are skipped.
* If number of threads per core is specified first, then the number of sockets and cores per socket are defaulted to all sockets and all cores per socket respectively.
* The number of cores per socket cannot be specified before sockets or after threads per core.
* The number of threads per core can be specified before or after core-offset (old syntax required it to be before core-offset);
* Parameters delimiter can be: empty, comma, lower-case x;
* Spaces are allowed around numbers, around letters, around delimiter.
Approximate shorthand specification:
KMP_PLACE_THREADS="[num_sockets(S|s)[[delim]offset(O|o)][delim]][num_cores_per_socket(C|c)[[delim]offset(O|o)][delim]][num_threads_per_core(T|t)]"
Differential Revision: http://reviews.llvm.org/D13175
llvm-svn: 249708
This patch adjusts the buffer size when reducing the buffer used for printing.
This solves the memory corruption in Windows debug library, and potential
memory corruption in other builds.
llvm-svn: 248588
This is a follow up to the hierarchy cleanup patch.
Added some clarifying comments to hierarchy_info.
Fixed a bug with the depth field not being updated cleanly during a resize.
Fixed resize to first check capacity as determined by maxLevels before actually doing the full resize.
Differential Revision: http://reviews.llvm.org/D12562
llvm-svn: 247333
Some of this is improvement to code suggested by Hal Finkel. Four changes here:
1.Cleanup of hierarchy code to handle all hierarchy cases whether affinity is available or not
2.Separated this and other classes and common functions out to a header file
3.Added a destructor-like fini function for the hierarchy (and call in __kmp_cleanup)
4.Remove some redundant code that is hopefully no longer needed
Differential Revision: http://reviews.llvm.org/D12449
llvm-svn: 247326
This patch fixes a bug when eliminating layers in the machine topology (namely
cores, and threads). Before this patch, if a user specifies using only one
thread per socket, then affinity is not set properly due to bad topology
pruning.
Differential Revision: http://reviews.llvm.org/D11158
llvm-svn: 245966
This fix allows the machine hierarchy to be expanded in case it needs to handle
more threads. It adds a resize function to accomplish this.
Differential Revision: http://reviews.llvm.org/D9900
llvm-svn: 240292
I tried to compile with Visual Studio using CMake and found these two sections of code
causing problems for Visual Studio. The first one removes the use of variable length
arrays by instead using KMP_ALLOCA(). The second part eliminates a redundant cpuid
assembly call by using the already existing __kmp_x86_cpuid() call instead.
llvm-svn: 240290
These changes are mostly in comments, but there are a few
that aren't. Change libiomp5 => libomp everywhere. One internal
function name is changed in kmp_gsupport.c, and in kmp_i18n.c, the
static char[] variable 'name' is changed to "libomp".
llvm-svn: 238712
understand that this is not friendly, and are working to change our
internal code-development to make it easier to make development
features available more frequently and in finer (more functional)
chunks. Unfortunately we haven't got that in place yet, and unpicking
this into multiple separate check-ins would be non-trivial, so please
bear with me on this one. We should be better in the future.
Apologies over, what do we have here?
GGC 4.9 compatibility
--------------------
* We have implemented the new entrypoints used by code compiled by GCC
4.9 to implement the same functionality in gcc 4.8. Therefore code
compiled with gcc 4.9 that used to work will continue to do so.
However, there are some other new entrypoints (associated with task
cancellation) which are not implemented. Therefore user code compiled
by gcc 4.9 that uses these new features will not link against the LLVM
runtime. (It remains unclear how to handle those entrypoints, since
the GCC interface has potentially unpleasant performance implications
for join barriers even when cancellation is not used)
--- new parallel entry points ---
new entry points that aren't OpenMP 4.0 related
These are implemented fully :-
GOMP_parallel_loop_dynamic()
GOMP_parallel_loop_guided()
GOMP_parallel_loop_runtime()
GOMP_parallel_loop_static()
GOMP_parallel_sections()
GOMP_parallel()
--- cancellation entry points ---
Currently, these only give a runtime error if OMP_CANCELLATION is true
because our plain barriers don't check for cancellation while waiting
GOMP_barrier_cancel()
GOMP_cancel()
GOMP_cancellation_point()
GOMP_loop_end_cancel()
GOMP_sections_end_cancel()
--- taskgroup entry points ---
These are implemented fully.
GOMP_taskgroup_start()
GOMP_taskgroup_end()
--- target entry points ---
These are empty (as they are in libgomp)
GOMP_target()
GOMP_target_data()
GOMP_target_end_data()
GOMP_target_update()
GOMP_teams()
Improvements in Barriers and Fork/Join
--------------------------------------
* Barrier and fork/join code is now in its own file (which makes it
easier to understand and modify).
* Wait/release code is now templated and in its own file; suspend/resume code is also templated
* There's a new, hierarchical, barrier, which exploits the
cache-hierarchy of the Intel(r) Xeon Phi(tm) coprocessor to improve
fork/join and barrier performance.
***BEWARE*** the new source files have *not* been added to the legacy
Cmake build system. If you want to use that fixes wil be required.
Statistics Collection Code
--------------------------
* New code has been added to collect application statistics (if this
is enabled at library compile time; by default it is not). The
statistics code itself is generally useful, the lightweight timing
code uses the X86 rdtsc instruction, so will require changes for other
architectures.
The intent of this code is not for users to tune their codes but
rather
1) For timing code-paths inside the runtime
2) For gathering general properties of OpenMP codes to focus attention
on which OpenMP features are most used.
Nested Hot Teams
----------------
* The runtime now maintains more state to reduce the overhead of
creating and destroying inner parallel teams. This improves the
performance of code that repeatedly uses nested parallelism with the
same resource allocation. Set the new KMP_HOT_TEAMS_MAX_LEVEL
envirable to a depth to enable this (and, of course, OMP_NESTED=true
to enable nested parallelism at all).
Improved Intel(r) VTune(Tm) Amplifier support
---------------------------------------------
* The runtime provides additional information to Vtune via the
itt_notify interface to allow it to display better OpenMP specific
analyses of load-imbalance.
Support for OpenMP Composite Statements
---------------------------------------
* Implement new entrypoints required by some of the OpenMP 4.1
composite statements.
Improved ifdefs
---------------
* More separation of concepts ("Does this platform do X?") from
platforms ("Are we compiling for platform Y?"), which should simplify
future porting.
ScaleMP* contribution
---------------------
Stack padding to improve the performance in their environment where
cross-node coherency is managed at the page level.
Redesign of wait and release code
---------------------------------
The code is simplified and performance improved.
Bug Fixes
---------
*Fixes for Windows multiple processor groups.
*Fix Fortran module build on Linux: offload attribute added.
*Fix entry names for distribute-parallel-loop construct to be consistent with the compiler codegen.
*Fix an inconsistent error message for KMP_PLACE_THREADS environment variable.
llvm-svn: 219214
Port the OpenMP runtime to FreeBSD along with associated build system changes.
Also begin to generalize affinity capabilities so they aren't tied explicitly
to Windows and Linux.
The port builds with stock clang and gmake and has no additional runtime
dependencies.
All but a handful of the validation suite tests are now passing on FreeBSD 10
x86_64.
llvm-svn: 202478
This release use aligns with Intel(r) Composer XE 2013 SP1 Product Update 2
New features
* The library can now be built with clang (though wiht some
limitations since clang does not support 128 bit floats)
* Support for Vtune analysis of load imbalance
* Code contribution from Steven Noonan to build the runtime for ARM*
architecture processors
* First implementation of runtime API for OpenMP cancellation
Bug Fixes
* Fixed hang on Windows (only) when using KMP_BLOCKTIME=0
llvm-svn: 197914
