This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295333
This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295319
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293005
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293001
The if-clause on the combined directive potentially applies to both the
'target' and the 'parallel' regions. Codegen'ing the if-clause on the
combined directive requires additional support because the expression in
the clause must be captured by the 'target' capture statement but not
the 'parallel' capture statement. Note that this situation arises for
other clauses such as num_threads.
The OMPIfClause class inherits OMPClauseWithPreInit to support capturing
of expressions in the clause. A member CaptureRegion is added to
OMPClauseWithPreInit to indicate which captured statement (in this case
'target' but not 'parallel') captures these expressions.
To ensure correct codegen of captured expressions in the presence of
combined 'target' directives, OMPParallelScope was added to 'parallel'
codegen.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28781
llvm-svn: 292437
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292419
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292374
This patch refactors code that calls codegen for target regions. Currently
the codebase only supports the 'target' directive. The patch pulls out
common target processing code into a static function that can be called
by codegen for any target directive.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28752
llvm-svn: 292134
This patch is to implement sema and parsing for 'target teams distribute simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28252
llvm-svn: 291579
https://reviews.llvm.org/D17840
This patch enables private, firstprivate, and lastprivate clauses for the OpenMP distribute directive.
Regression tests differ from the similar case of the same clauses on the for directive, by removing a reference to two global variables g and g1. This is necessary because: 1. a distribute pragma is only allowed inside a target region; 2. referring a global variable (e.g. g and g1) in a target region requires the program to enclose the variable in a "declare target" region; 3. declare target pragmas, which are used to define a declare target region, are currently unavailable in clang (patch being prepared).
For this reason, I moved the global declarations into local variables.
llvm-svn: 290898
This patch is to implement sema and parsing for 'target teams distribute parallel for simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28202
llvm-svn: 290862
This patch is to implement sema and parsing for 'target teams distribute parallel for’ pragma.
Differential Revision: https://reviews.llvm.org/D28160
llvm-svn: 290725
This patch is to implement sema and parsing for 'target teams distribute' pragma.
Differential Revision: https://reviews.llvm.org/D28015
llvm-svn: 290508
This patch is to implement sema and parsing for 'teams distribute parallel for' pragma.
Differential Revision: https://reviews.llvm.org/D27345
llvm-svn: 289179
This patch is to implement sema and parsing for 'teams distribute parallel for simd' pragma.
Differential Revision: https://reviews.llvm.org/D27084
llvm-svn: 288294
If 'omp cancel' construct is used in a worksharing construct it may
cause hanging of the software in case if reduction clause is used. Patch fixes this problem by avoiding extra reduction processing for branches that were canceled.
llvm-svn: 287227
Summary:
r286944 introduced bugs detected by ASAN as use-after-return.
r287025 have not fixed them completely.
This reverts commit r286944 and r287025.
Reviewers: ABataev
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D26720
llvm-svn: 287069
If 'omp cancel' construct is used in a worksharing construct it may cause
hanging of the software in case if reduction clause is used. Patch fixes
this problem by avoiding extra reduction processing for branches that
were canceled.
llvm-svn: 286944
can be used to improve the locations when generating remarks for loops.
Depends on the companion LLVM change r286227.
Patch by Florian Hahn.
Differential Revision: https://reviews.llvm.org/D25764
llvm-svn: 286456
After some changes in codegen capturing of VLA variables in OpenMP regions was broken, causing compiler crash. Patch fixes this issue.
llvm-svn: 286103
After some changes in codegen capturing of VLA variables in OpenMP
regions was broken, causing compiler crash. Patch fixes this issue.
llvm-svn: 286098
constexpr variable.
When compiling a constexpr NSString initialized with an objective-c
string literal, CodeGen emits objc_storeStrong on an uninitialized
alloca, which causes a crash.
This patch folds the code in EmitScalarInit into EmitStoreThroughLValue
and fixes the crash by calling objc_retain on the string instead of
using objc_storeStrong.
rdar://problem/28562009
Differential Revision: https://reviews.llvm.org/D25547
llvm-svn: 284516
access, by Erich Keane
OpenMP creates a variable array type with a a null size-expr. The Debug
generation failed to due to this. This patch corrects the openmp
implementation, updates the tests, and adds a new one for this
condition.
Differential Revision: https://reviews.llvm.org/D25373
llvm-svn: 284110
This reverts commit r279003 as it breaks some of our buildbots (e.g.
clang-cmake-aarch64-quick, clang-x86_64-linux-selfhost-modules).
The error is in OpenMP/teams_distribute_simd_ast_print.cpp:
clang: /home/buildslave/buildslave/clang-cmake-aarch64-quick/llvm/include/llvm/ADT/DenseMap.h:527:
bool llvm::DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT>::LookupBucketFor(const LookupKeyT&, const BucketT*&) const
[with LookupKeyT = clang::Stmt*; DerivedT = llvm::DenseMap<clang::Stmt*, long unsigned int>;
KeyT = clang::Stmt*; ValueT = long unsigned int;
KeyInfoT = llvm::DenseMapInfo<clang::Stmt*>;
BucketT = llvm::detail::DenseMapPair<clang::Stmt*, long unsigned int>]:
Assertion `!KeyInfoT::isEqual(Val, EmptyKey) && !KeyInfoT::isEqual(Val, TombstoneKey) &&
"Empty/Tombstone value shouldn't be inserted into map!"' failed.
llvm-svn: 279045
This patch is to implement sema and parsing for 'teams distribute simd’ pragma.
This patch is originated by Carlo Bertolli.
Differential Revision: https://reviews.llvm.org/D23528
llvm-svn: 279003
Summary: This patch adds support for the use_device_ptr clause. It includes changes in SEMA that could not be tested without codegen, namely, the use of the first private logic and mappable expressions support.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev
Subscribers: caomhin, cfe-commits
Differential Revision: https://reviews.llvm.org/D22691
llvm-svn: 276977
Summary:
This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash.
Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region.
Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev
Subscribers: caomhin, cfe-commits
Differential Revision: https://reviews.llvm.org/D22690
llvm-svn: 276933
This patch is to implement sema and parsing for 'target parallel for simd' pragma.
Differential Revision: http://reviews.llvm.org/D22096
llvm-svn: 275365
http://reviews.llvm.org/D21904
This patch is similar to the implementation of 'private' clause: it adds a list of private pointers to be used within the target data region to store the device pointers returned by the runtime.
Please refer to the following document for a full description of what the runtime witll return in this case (page 10 and 11):
https://github.com/clang-omp/OffloadingDesign
I am happy to answer any question related to the runtime interface to help reviewing this patch.
llvm-svn: 275271
Summary: This patch is an implementation of sema and parsing for the OpenMP composite pragma 'distribute simd'.
Differential Revision: http://reviews.llvm.org/D22007
llvm-svn: 274604
Summary: This patch is an implementation of sema and parsing for the OpenMP composite pragma 'distribute parallel for simd'.
Differential Revision: http://reviews.llvm.org/D21977
llvm-svn: 274530
[OpenMP] Initial implementation of parse and sema for composite pragma 'distribute parallel for'
This patch is an initial implementation for #distribute parallel for.
The main differences that affect other pragmas are:
The implementation of 'distribute parallel for' requires blocking of the associated loop, where blocks are "distributed" to different teams and iterations within each block are scheduled to parallel threads within each team. To implement blocking, sema creates two additional worksharing directive fields that are used to pass the team assigned block lower and upper bounds through the outlined function resulting from 'parallel'. In this way, scheduling for 'for' to threads can use those bounds.
As a consequence of blocking, the stride of 'distribute' is not 1 but it is equal to the blocking size. This is returned by the runtime and sema prepares a DistIncrExpr variable to hold that value.
As a consequence of blocking, the global upper bound (EnsureUpperBound) expression of the 'for' is not the original loop upper bound (e.g. in for(i = 0 ; i < N; i++) this is 'N') but it is the team-assigned block upper bound. Sema creates a new expression holding the calculation of the actual upper bound for 'for' as UB = min(UB, PrevUB), where UB is the loop upper bound, and PrevUB is the team-assigned block upper bound.
llvm-svn: 273884
http://reviews.llvm.org/D21564
This patch is an initial implementation for #distribute parallel for.
The main differences that affect other pragmas are:
The implementation of 'distribute parallel for' requires blocking of the associated loop, where blocks are "distributed" to different teams and iterations within each block are scheduled to parallel threads within each team. To implement blocking, sema creates two additional worksharing directive fields that are used to pass the team assigned block lower and upper bounds through the outlined function resulting from 'parallel'. In this way, scheduling for 'for' to threads can use those bounds.
As a consequence of blocking, the stride of 'distribute' is not 1 but it is equal to the blocking size. This is returned by the runtime and sema prepares a DistIncrExpr variable to hold that value.
As a consequence of blocking, the global upper bound (EnsureUpperBound) expression of the 'for' is not the original loop upper bound (e.g. in for(i = 0 ; i < N; i++) this is 'N') but it is the team-assigned block upper bound. Sema creates a new expression holding the calculation of the actual upper bound for 'for' as UB = min(UB, PrevUB), where UB is the loop upper bound, and PrevUB is the team-assigned block upper bound.
llvm-svn: 273705