Summary:
In non-SPMD mode we may end up with the divergent threads when trying to
increment/decrement parallel level counter. It may lead to incorrect
calculations of the parallel level and wrong results when threads are
divergent. We need to reconverge the threads before trying to modify the
parallel level counter.
Reviewers: grokos, jdoerfert
Subscribers: guansong, openmp-commits, caomhin, kkwli0
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66802
llvm-svn: 370803
Summary:
Use target_impl functions to replace more inline asm
Follow on from D65836. Removes remaining asm shuffles and lanemask accessors
Also changes the types of target_impl bitwise functions to unsigned.
Reviewers: jdoerfert, ABataev, grokos, Hahnfeld, gregrodgers, ronlieb, hfinkel
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66809
llvm-svn: 370216
Summary:
[libomptarget] Refactor syncthreads macro to inline function
See also abandoned D66846, split into this diff and others.
Rev 2 of D66855
Reviewers: jdoerfert, ABataev, grokos, ronlieb, gregrodgers
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66861
llvm-svn: 370210
Summary:
[libomptarget] Refactor syncwarp macro to inline function
See also abandoned D66846, split into this diff and others.
Reviewers: jdoerfert, ABataev, grokos, ronlieb, gregrodgers
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66857
llvm-svn: 370149
Summary:
[libomptarget] Refactor shfl_down_sync macro to inline function
See also abandoned D66846, split into this diff and others.
Reviewers: jdoerfert, ABataev, grokos, ronlieb, gregrodgers
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66853
llvm-svn: 370146
Summary:
[libomptarget] Refactor shfl_sync macro to inline function
See also abandoned D66846, split into this diff and others.
Reviewers: jdoerfert, ABataev, grokos, ronlieb, gregrodgers
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66852
llvm-svn: 370144
Summary:
Added function void __kmpc_syncwarp(int32_t) to expose it to the
compiler. It is required to fix the problem with the critical regions in
Cuda9.0+. We cannot use barrier in the critical region, but still need
to reconverge the threads in the warp after. This function allows to do
this.
Reviewers: grokos, jdoerfert
Subscribers: guansong, openmp-commits, kkwli0, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D66672
llvm-svn: 369933
Summary:
In Cuda 9.0 it is not guaranteed that threads in the warps are
convergent. We need to use __syncwarp() function to reconverge
the threads and to guarantee the memory ordering among threads in the
warps.
This is the first patch to fix the problem with the test
libomptarget/deviceRTLs/nvptx/src/sync.cu on Cuda9+.
This patch just replaces calls to __shfl_sync() function with the call
of __syncwarp() function where we need to reconverge the threads when we
try to modify the value of the parallel level counter.
Reviewers: grokos
Subscribers: guansong, jfb, jdoerfert, caomhin, kkwli0, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65013
llvm-svn: 369796
Summary:
[libomptarget] Factor architecture dependent code out of loop.cu
Related to the patch series starting D64217. Added subscribers to said series as reviewers. This effort is smaller in scope.
This patch factors out just enough architecture dependent code from loop.cu to allow the same source to be used with amdgcn, given a different target_impl.h. Testing is that the same bitcode (modulo variable names) is generated for libomptarget before and after the refactor, for nvptx and the out of tree amdgcn.
Reviewers: jdoerfert, ABataev, bollu, jfb, tra, grokos, Hahnfeld, guansong, xtian, gregrodgers, ronlieb, hfinkel, gtbercea, guraypp, arpith-jacob
Reviewed By: jdoerfert, ABataev
Subscribers: dexonsmith, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65836
llvm-svn: 368751
Summary:
This patch adds support for the close map modifier.
The close map modifier will overwrite the unified shared memory requirement and create a device copy of the data.
Reviewers: ABataev, Hahnfeld, caomhin, grokos, jdoerfert, AlexEichenberger
Reviewed By: Hahnfeld, AlexEichenberger
Subscribers: guansong, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65340
llvm-svn: 368488
We have one global RTLs.RequiresFlags, I don't see a need to make a
copy per device that the runtime manages. This was problematic anyway
because the copy happened during the first __tgt_register_lib(). This
made it impossible to call __tgt_register_requires() from normal user
funtions for testing.
Hence, this change also fixes unified_shared_memory/shared_update.c for
older versions of Clang that don't call __tgt_register_requires() before
__tgt_register_lib().
Differential Revision: https://reviews.llvm.org/D66019
llvm-svn: 368465
Summary:
This patch adds support for using unified memory in the case of regular maps that happen when a target region is offloaded to the device.
For cases where only a single version of the data is required then the host address can be used. When variables need to be privatized in any way or globalized, then the copy to the device is still required for correctness.
Reviewers: ABataev, jdoerfert, Hahnfeld, AlexEichenberger, caomhin, grokos
Reviewed By: Hahnfeld
Subscribers: mgorny, guansong, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65001
llvm-svn: 368192
Summary:
[libomptarget] Use forceinline. Necessary for nvcc to inline small functions within the bitcode library
Suggested in D65836
Reviewers: ABataev, jdoerfert, grokos, gregrodgers
Subscribers: openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65876
llvm-svn: 368177
Ensures that CUDA fail reasons (such as "No CUDA-capable device detected")
are printed together with libomptarget's debug message
(e.g. "Error when setting CUDA context"). Previously, the former was
printed only in CMAKE_BUILD_TYPE=Debug builds while the latter was
enabled by LIBOMPTARGET_ENABLE_DEBUG.
With this change, also only call cuGetErrorString when the error will be
printed.
Suggested-by: Ye Luo <xw111luoye@gmail.com>
Differential Revision: https://reviews.llvm.org/D65687
llvm-svn: 367910
This patch implements the libomptarget runtime interface for OpenMP 5.0
declare mapper functions. The declare mapper functions generated by
Clang will call them to complete the mapping of members.
kmpc_mapper_num_components gets the current number of components for a
user-defined mapper; kmpc_push_mapper_component pushes back one
component for a user-defined mapper.
The design slides can be found at
https://github.com/lingda-li/public-sharing/blob/master/mapper_runtime_design.pptx
Patch by Lingda Li <lildmh@gmail.com>
Differential Revision: https://reviews.llvm.org/D60972
llvm-svn: 367772
Summary:
According to the OpenMP standard, barrier operation must perform
implicit flush operation. Currently, if there is only one thread in the
team, barrier does not flush the memory. Patch fixes this problem.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D62398
llvm-svn: 367024
If the first target region in a program calls the push_tripcount
function, libomptarget didn't handle the offload policy correctly.
This could lead to unexpected error messages as seen in
http://lists.llvm.org/pipermail/openmp-dev/2019-June/002561.html
To solve this, add a check calling IsOffloadDisabled() as all other
entry points already do. If this method returns false, libomptarget
is effectively disabled.
Differential Revision: https://reviews.llvm.org/D64626
llvm-svn: 366810
Remove loopTripCnt from threaded device stack after consuming it.
Added a libomptarget DP message to aid in future debugging and to
validate the added testcase, which only runs in Debug build.
Differential Revision: https://reviews.llvm.org/D64808
llvm-svn: 366349
Summary:
We used CUDART_VERSION macro to check for the installed cuda version
but this macro is defined in cuda_runtime_api.h, which is not used by
project. Better to use CUDA_VERSION macro, which is defined in cuda.h.
Also, added the check if this macro is defined. If macro is undefined,
there is something wrong with the cuda configuration and we should not
continue the compilation.
This also fixes problems with runtime building in cuda 10+.
Reviewers: grokos
Subscribers: guansong, jdoerfert, caomhin, kkwli0, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D64648
llvm-svn: 366224
Summary:
We used to call __kmpc_omp_taskwait function with global threadid set to
0. It may crash the application at the runtime if the thread executing
target region is not a master thread.
Reviewers: grokos, kkwli0
Subscribers: guansong, jdoerfert, caomhin, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D64571
llvm-svn: 366220
These entry points are never called by Clang trunk nor clang-ykt. If
XL doesn't use them either, they can finally go away.
Differential Revision: https://reviews.llvm.org/D52700
llvm-svn: 365817
Summary:
__kmpc_push_tripcount function is not thread safe and may lead to data
race when the target regions are executed in parallel threads. The patch
makes loopTripCnt counter thread aware and stores the tripcount value
per thread in the map. Access to map is guarded by mutex to prevent
data race in the map itself.
Test is for NVPTX target because it does not work correctly on the
host. Seems to me, there is a problem in libomp with target regions in
the parallel threads.
Reviewers: grokos
Subscribers: guansong, jfb, jdoerfert, openmp-commits, kkwli0, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D64080
llvm-svn: 365332
Summary:
According to the OpenMP standard, flush makes a thread’s temporary view of memory consistent with memory and enforces an order on the memory operations of the variables explicitly specified or implied.
According to the Cuda toolkit documentation (https://docs.nvidia.com/cuda/archive/8.0/cuda-c-programming-guide/index.html#memory-fence-functions), __threadfence() functions provides required functionality.
__threadfence_system() also provides required functionality, but it also
includes some extra functionality, like synchronization of page-locked
host memory, synchronization for the host, etc. It is not required per
the standard and we can use more relaxed version of memory fence
operation.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jfb, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D62397
llvm-svn: 364572
Summary:
This patch adds support for handling variables under the:
```
#pragma omp declare target to()
```
clause when the
```
#pragma omp requires unified_shared_memory
```
is used.
The address of the host variable is copied into the device pointer just like for the declare target link case.
Reviewers: ABataev, caomhin, grokos, AlexEichenberger
Reviewed By: grokos
Subscribers: jcownie, guansong, jdoerfert, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D63106
llvm-svn: 363825
Summary:
The problems with __syncthreads() were fixed in clang >= 9.0 and the
original __syncthreads() can be used instead of the ptx instruction.
Reviewers: grokos
Subscribers: guansong, jdoerfert, openmp-commits, kkwli0, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D63515
llvm-svn: 363807
Summary: This patch enables the usage of a host variable on the device for declare target link variables when unified memory is available.
Reviewers: ABataev, caomhin, grokos
Reviewed By: grokos
Subscribers: Hahnfeld, guansong, jdoerfert, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D60884
llvm-svn: 362505
Summary:
Parallel level counter should be volatile to prevent some dangerous
optimiations by the ptxas. Otherwise, ptxas optimizations lead to
undefined behaviour in some cases.
Also, use __threadfence() for #pragma omp flush and if the barrier
should not be used (we have only one thread in the team), still perform
flush operation since the standard requires implicit flush when
executing barriers.
Reviewers: gtbercea, kkwli0, grokos
Subscribers: guansong, jfb, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D62199
llvm-svn: 361421
Summary:
Target link variables are currently implemented by creating a copy of the variables on the device side and unified memory never gets exploited.
When the prgram uses the:
```
#pragma omp requires unified_shared_memory
```
directive in conjunction with a declare target link, the linked variable is no longer allocated on the device and the host version is used instead.
This behavior is overridden by performing an explicit mapping.
A Clang side patch is required.
Reviewers: ABataev, AlexEichenberger, grokos, Hahnfeld
Reviewed By: AlexEichenberger, grokos, Hahnfeld
Subscribers: Hahnfeld, jfb, guansong, jdoerfert, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D60223
llvm-svn: 361294
Summary:
Patch improves performance of the full runtime mode by moving
threads limit counter to the shared memory. It also allows to save
global memory.
Reviewers: grokos, kkwli0, gtbercea
Subscribers: guansong, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61801
llvm-svn: 360584
Summary:
Patch improves performance of the full runtime mode by moving
number-of-threads counter to the shared memory. It also allows to save
global memory.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jfb, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61785
llvm-svn: 360457
Summary:
Patch improves performance of the full runtime mode by moving
thread-limit counter to the shared memory. It also allows to save
global memory.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jdoerfert, caomhin, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61526
llvm-svn: 359922
Summary:
Used parallelLevel[] counter to simplify and improve implementation of
the existing standard OpenMP functions. Functions are tested already in
several tests, the patch is NFC.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jdoerfert, caomhin, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61459
llvm-svn: 359892
Summary:
Previously for the different purposes we need to get the active/common
parallel level and with full runtime we iterated over all the records to
calculate this level. Instead, we can used the warp-based parallel level
counters used in no-runtime mode.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jfb, jdoerfert, caomhin, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61395
llvm-svn: 359822
Summary:
Function omp_get_thread_limit() in SPMD mode can return the maximum
available number of threads as a result.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D61378
llvm-svn: 359790
Summary:
The parallelLevel counter must be on per-thread basis to fully support
L2+ parallelism, otherwise we may end up with undefined behavior.
Introduce the parallelLevel on per-warp basis using shared memory. It
allows to avoid the problems with the synchronization and allows fully
support L2+ parallelism in SPMD mode with no runtime.
Reviewers: gtbercea, grokos
Subscribers: guansong, jdoerfert, caomhin, kkwli0, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D60918
llvm-svn: 359341
Fix the test to run it really in SPMD mode without runtime. Previously
it was run in SPMD + full runtime mode and does not allow to cehck the
functionality correctly.
llvm-svn: 358902
Summary:
If the kernel is executed in SPMD mode and the L2+ parallel for region
with the dynamic scheduling is executed, dynamic scheduling functions
are called. They expect full runtime support, but SPMD kernels may be
executed without the full runtime. It leads to the runtime crash of the
compiled program. Patch fixes this problem + fixes handling of the
parallelism level in SPMD mode, which is required as part of this patch.
Reviewers: gtbercea, kkwli0, grokos
Subscribers: guansong, jdoerfert, openmp-commits, caomhin
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D60578
llvm-svn: 358442
At the moment, support for runtime debug output using the
OMPTARGET_DEBUG=1 environment variable is only available with
CMAKE_BUILD_TYPE=Debug builds. The patch allows setting it independently
using the LIBOMPTARGET_ENABLE_DEBUG option, which is enabled by default
depending on CMAKE_BUILD_TYPE. That is, unless this option is set
explicitly, nothing changes. This is the same mechanism used by LLVM for
LLVM_ENABLE_ASSERTIONS.
This patch also removes adding -g -O0 in debug builds, it should be
handled by cmake's CMAKE_{C|CXX}_FLAGS_DEBUG configuration option.
Idea by Hal Finkel
Differential Revision: https://reviews.llvm.org/D55952
llvm-svn: 356998
Summary:
This patch adds a more sophisticated team reduction scheme to the OpenMP libomptarget-nvptx runtime.
The scheme uses a fixed size global memory buffer whose length can be adjusted via compiler flag:
```
-fopenmp-cuda-teams-reduction-recs-num=1024
```
The global buffer is a structure of arrays (with default size of 1024 each and controlled by the above flag), one array for each reduction variable.
Values in the buffer are processed by the last team to finish executing the body of the target region.
In addition to adding support for the new flag, the compiler also emits special functions used for the reduction of the intermediate reduction values. These changes will be added in a separate compiler patch following this one.
Reviewers: ABataev, caomhin
Reviewed By: ABataev
Subscribers: guansong, jfb, jdoerfert, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D58409
llvm-svn: 354471
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
Summary: Replace existing infrastructure for tracking parallel level using global memory with a per-team shared memory variable. This minimizes the impact of the overhead of tracking the parallel level for non-nested cases.
Reviewers: ABataev, caomhin
Reviewed By: ABataev
Subscribers: guansong, openmp-commits
Differential Revision: https://reviews.llvm.org/D55773
llvm-svn: 350747
Summary:
Previous implementation may cause the runtime crash when the number of
teams is > 1024. Patch fixes this problem + reduces number of the atomic
operations by 32 times.
Reviewers: grokos, gtbercea, kkwli0
Subscribers: guansong, jfb, openmp-commits, caomhin
Differential Revision: https://reviews.llvm.org/D56332
llvm-svn: 350524
