if the function has globalized variables and called in context of
target/teams/distribute regions, it does not need to globalize 32
copies of the same variables for memory coalescing, it is enough to
have just one copy, because there is parallel region.
Patch does this by adding call for `__kmpc_parallel_level` function and
checking its return value. If the code sees that the parallel level is
0, then only one variable is allocated, not 32.
llvm-svn: 344356
target/teams/distribute regions.
Previously introduced globalization scheme that uses memory coalescing
scheme may increase memory usage fr the variables that are devlared in
target/teams/distribute contexts. We don't need 32 copies of such
variables, just 1. Patch reduces memory use in this case.
llvm-svn: 344273
Added support for memory coalescing for better performance for
globalized variables. From now on all the globalized variables are
represented as arrays of 32 elements and each thread accesses these
elements using `tid & 31` as index.
llvm-svn: 344049
Fixed emission of the __kmpc_global_thread_num() so that it is not
messed up with alloca instructions anymore. Plus, fixes emission of the
__kmpc_global_thread_num() functions in the target outlined regions so
that they are not called before runtime is initialized.
llvm-svn: 343856
Worker threads fork off to the compiler generated worker function
directly after entering the kernel function. Hence, there is no
need to check whether the current thread is the master if we are
outside of a parallel region (neither SPMD nor parallel_level > 0).
Differential Revision: https://reviews.llvm.org/D52732
llvm-svn: 343618
lightweight runtime.
The datasharing flag must be set to `1` when executing SPMD-mode compatible directive with reduction|lastprivate clauses.
llvm-svn: 343492
Summary: Set default schedule for parallel for loops to schedule(static, 1) when using SPMD mode on the NVPTX device offloading toolchain to ensure coalescing.
Reviewers: ABataev, Hahnfeld, caomhin
Reviewed By: ABataev
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D52629
llvm-svn: 343260
Summary: For the OpenMP NVPTX toolchain choose a default distribute schedule that ensures coalescing on the GPU when in SPMD mode. This significantly increases the performance of offloaded target code and reduces the number of registers used on the GPU side.
Reviewers: ABataev, caomhin, Hahnfeld
Reviewed By: ABataev, Hahnfeld
Subscribers: Hahnfeld, jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D52434
llvm-svn: 343253
Add support for OMP5.0 requires directive and unified_address clause.
Patches to follow will include support for additional clauses.
Differential Revision: https://reviews.llvm.org/D52359
llvm-svn: 343063
'declare target'.
All the functions, referenced in implicit|explicit target regions must
be emitted during code emission for the device.
llvm-svn: 341093
If the target construct can be executed in SPMD mode + it is a loop
based directive with static scheduling, we can use lightweight runtime
support.
llvm-svn: 340953
The first argument for the parallel outlined functions, called as
serialized parallel regions, should be a pointer to the global thread id
that always is 0.
llvm-svn: 337957
Summary: In the SPMD case, we need to initialize the data sharing and globalization infrastructure. This covers the case when an SPMD region calls a function in a different compilation unit.
Reviewers: ABataev, carlo.bertolli, caomhin
Reviewed By: ABataev
Subscribers: Hahnfeld, jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D49188
llvm-svn: 337015
In generic data-sharing mode we are allowed to not globalize local
variables that escape their declaration context iff they are declared
inside of the parallel region. We can do this because L2 parallel
regions are executed sequentially and, thus, we do not need to put
shared local variables in the global memory.
llvm-svn: 336567
Patch tries to make better analysis of the variables that should be
globalized. From now, instead of all parallel directives it will check
only distribute parallel .. directives and check only for
firstprivte/lastprivate variables if they must be globalized.
llvm-svn: 335632
If the shuffle is required for the reduced structures/big data type,
current code may cause compiler crash because of the loading of the
aggregate values. Patch fixes this problem.
llvm-svn: 335377
parallel region.
If the current construct requires sharing of the local variable in the
inner parallel region, this variable must be globalized to avoid
runtime crash.
llvm-svn: 335285
If orphaned parallel region is found, the next code must be emitted:
```
if(__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid))
Serialized execution.
else if (IsMasterThread())
Prepare and signal worker.
else
Outined function call.
```
llvm-svn: 333301
If the orphaned directive is executed in SPMD mode, we need to emit the
check for the SPMD mode and run the orphaned parallel directive in
sequential mode.
llvm-svn: 332467
In generic data-sharing mode we do not need to globalize
variables/parameters of reference/pointer types. They already are placed
in the global memory.
llvm-svn: 332380
Added initial support for L2 parallelism in SPMD mode. Note, though,
that the orphaned parallel directives are not currently supported in
SPMD mode.
llvm-svn: 332016
This is similar to the LLVM change https://reviews.llvm.org/D46290.
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\@brief'); do perl -pi -e 's/\@brief //g' $i & done
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46320
llvm-svn: 331834
Added initial codegen for level 2, 3 etc. parallelism. Currently, all
the second, the third etc. parallel regions will run sequentially.
llvm-svn: 331642
Some symbols are not allowed to be used as names on some targets. Patch
ries to unify the emission of the names of LLVM globals so they could be
used on different targets.
llvm-svn: 331358
NVPTX target.
When generating the wrapper function for the offloading region, we need
to call the outlined function and cast the arguments correctly to follow
the ABI. Usually, variables captured by value are casted to `uintptr_t`
type. But this should not performed for the variables with pointer type.
llvm-svn: 330620
Added NUW flags for all the add|mul|sub operations + replaced sdiv by udiv
as we operate on unsigned values only (addresses, converted to integers)
llvm-svn: 329411
variables.
Added emission of the offloading data sections for the variables within
declare target regions + fixes emission of the declare target variables
marked as declare target not within the declare target region.
llvm-svn: 328888
Summary: The workers also need to initialize the global stack. The call to the initialization function needs to happen after the kernel_init() function is called by the master. This ensures that the per-team data structures of the runtime have been initialized.
Reviewers: ABataev, grokos, carlo.bertolli, caomhin
Reviewed By: ABataev
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D44749
llvm-svn: 328219
constructs in generic mode.
Fixed codegen for distribute parallel combined constructs. We have to
pass and read the shared lower and upper bound from the distribute
region in the inner parallel region. Patch is for generic mode.
llvm-svn: 327990
If the generic codegen is enabled and private copy of the original
variable escapes the declaration context, this private copy should be
globalized just like it was the original variable.
llvm-svn: 327985
The compiler complained about
../tools/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp:184:15: error: unused variable 'CSI' [-Werror,-Wunused-variable]
if (auto *CSI = CGF.CapturedStmtInfo) {
^
1 error generated.
I don't know this code but it seems like an easy fix so I push it anyway
to get rid of the warning.
llvm-svn: 327694
If the variable is captured by value and the corresponding parameter in
the outlined function escapes its declaration context, this parameter
must be globalized. To globalize it we need to get the address of the
original parameter, load the value, store it to the global address and
use this global address instead of the original.
Patch improves globalization for parallel|teams regions + functions in
declare target regions.
llvm-svn: 327654
Summary: Remove this scheme for now since it will be covered by another more generic scheme using global memory. This code will be worked into an optimization for the generic data sharing scheme. Removing this completely and then adding it via future patches will make all future data sharing patches cleaner.
Reviewers: ABataev, carlo.bertolli, caomhin
Reviewed By: ABataev
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D43625
llvm-svn: 326948
Differential Revision: https://reviews.llvm.org/D43852
This patch extends the SPMD implementation to all target constructs and guards this implementation under a new flag.
llvm-svn: 326368
Summary:
This patch enables debugging of C99 VLA types by generating more precise
LLVM Debug metadata, using the extended DISubrange 'count' field that
takes a DIVariable.
This should implement:
Bug 30553: Debug info generated for arrays is not what GDB expects (not as good as GCC's)
https://bugs.llvm.org/show_bug.cgi?id=30553
Reviewers: echristo, aprantl, dexonsmith, clayborg, pcc, kristof.beyls, dblaikie
Reviewed By: aprantl
Subscribers: jholewinski, schweitz, davide, fhahn, JDevlieghere, cfe-commits
Differential Revision: https://reviews.llvm.org/D41698
llvm-svn: 323952
getAssociatedStmt() returns the outermost captured statement for the
OpenMP directive. It may return incorrect region in case of combined
constructs. Reworked the code to reduce the number of calls of
getAssociatedStmt() and used getInnermostCapturedStmt() and
getCapturedStmt() functions instead.
In case of firstprivate variables it may lead to an extra allocas
generation for private copies even if the variable is passed by value
into outlined function and could be used directly as private copy.
llvm-svn: 322393
If the reduction required shuffle in the NVPTX codegen, we may need to
cast the reduced value to the integer type. This casting was implemented
incorrectly and may cause compiler crash. Patch fixes this problem.
llvm-svn: 321818
Summary:
The backend should only emit data sharing code for the cases where it is needed.
A new function attribute is used by Clang to enable data sharing only for the cases where OpenMP semantics require it and there are variables that need to be shared.
Reviewers: hfinkel, Hahnfeld, ABataev, carlo.bertolli, caomhin
Reviewed By: ABataev
Subscribers: cfe-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D41123
llvm-svn: 320527
In the future the compiler will analyze whether the OpenMP
runtime needs to be (fully) initialized and avoid that overhead
if possible. The functions already take an argument to transfer
that information to the runtime, so pass in the default value 1.
(This is needed for binary compatibility with libomptarget-nvptx
currently being upstreamed.)
Differential Revision: https://reviews.llvm.org/D40354
llvm-svn: 318836
Summary:
This patch is part of the development effort to add support in the current OpenMP GPU offloading implementation for implicitly sharing variables between a target region executed by the team master thread and the worker threads within that team.
This patch is the first of three required for successfully performing the implicit sharing of master thread variables with the worker threads within a team. The remaining two patches are:
- Patch D38978 to the LLVM NVPTX backend which ensures the lowering of shared variables to an device memory which allows the sharing of references;
- Patch (coming soon) is a patch to libomptarget runtime library which ensures that a list of references to shared variables is properly maintained.
A simple code snippet which illustrates an implicit data sharing situation is as follows:
```
#pragma omp target
{
// master thread only
int v;
#pragma omp parallel
{
// worker threads
// use v
}
}
```
Variable v is implicitly shared from the team master thread which executes the code in between the target and parallel directives. The worker threads must operate on the latest version of v, including any updates performed by the master.
The code generated in this patch relies on the LLVM NVPTX patch (mentioned above) which prevents v from being lowered in the thread local memory of the master thread thus making the reference to this variable un-shareable with the workers. This ensures that the code generated by this patch is correct.
Since the parallel region is outlined the passing of arguments to the outlined regions must preserve the original order of arguments. The runtime therefore maintains a list of references to shared variables thus ensuring their passing in the correct order. The passing of arguments to the outlined parallel function is performed in a separate function which the data sharing infrastructure constructs in this patch. The function is inlined when optimizations are enabled.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, Hahnfeld, ABataev, caomhin
Reviewed By: ABataev
Subscribers: cfe-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D38976
llvm-svn: 318773
Summary:
Convert clang::LangAS to a strongly typed enum
Currently both clang AST address spaces and target specific address spaces
are represented as unsigned which can lead to subtle errors if the wrong
type is passed. It is especially confusing in the CodeGen files as it is
not possible to see what kind of address space should be passed to a
function without looking at the implementation.
I originally made this change for our LLVM fork for the CHERI architecture
where we make extensive use of address spaces to differentiate between
capabilities and pointers. When merging the upstream changes I usually
run into some test failures or runtime crashes because the wrong kind of
address space is passed to a function. By converting the LangAS enum to a
C++11 we can catch these errors at compile time. Additionally, it is now
obvious from the function signature which kind of address space it expects.
I found the following errors while writing this patch:
- ItaniumRecordLayoutBuilder::LayoutField was passing a clang AST address
space to TargetInfo::getPointer{Width,Align}()
- TypePrinter::printAttributedAfter() prints the numeric value of the
clang AST address space instead of the target address space.
However, this code is not used so I kept the current behaviour
- initializeForBlockHeader() in CGBlocks.cpp was passing
LangAS::opencl_generic to TargetInfo::getPointer{Width,Align}()
- CodeGenFunction::EmitBlockLiteral() was passing a AST address space to
TargetInfo::getPointerWidth()
- CGOpenMPRuntimeNVPTX::translateParameter() passed a target address space
to Qualifiers::addAddressSpace()
- CGOpenMPRuntimeNVPTX::getParameterAddress() was using
llvm::Type::getPointerTo() with a AST address space
- clang_getAddressSpace() returns either a LangAS or a target address
space. As this is exposed to C I have kept the current behaviour and
added a comment stating that it is probably not correct.
Other than this the patch should not cause any functional changes.
Reviewers: yaxunl, pcc, bader
Reviewed By: yaxunl, bader
Subscribers: jlebar, jholewinski, nhaehnle, Anastasia, cfe-commits
Differential Revision: https://reviews.llvm.org/D38816
llvm-svn: 315871
When translating arguments for NVPTX target it is not taken into account
that function may have variable number of arguments. Patch fixes this
problem.
llvm-svn: 310920
Arguments, passed to the outlined function, must have correct address
space info for proper Debug info support. Patch sets global address
space for arguments that are mapped and passed by reference.
Also, cuda-gdb does not handle reference types correctly, so reference
arguments are represented as pointers.
llvm-svn: 310387
Arguments, passed to the outlined function, must have correct address
space info for proper Debug info support. Patch sets global address
space for arguments that are mapped and passed by reference.
Also, cuda-gdb does not handle reference types correctly, so reference
arguments are represented as pointers.
llvm-svn: 310377
Arguments, passed to the outlined function, must have correct address
space info for proper Debug info support. Patch sets global address
space for arguments that are mapped and passed by reference.
Also, cuda-gdb does not handle reference types correctly, so reference
arguments are represented as pointers.
llvm-svn: 310360
Arguments, passed to the outlined function, must have correct address
space info for proper Debug info support. Patch sets global address
space for arguments that are mapped and passed by reference.
Also, cuda-gdb does not handle reference types correctly, so reference
arguments are represented as pointers.
llvm-svn: 310104
Summary:
If the first parameter of the function is the ImplicitParamDecl, codegen
automatically marks it as an implicit argument with `this` or `self`
pointer. Added internal kind of the ImplicitParamDecl to separate
'this', 'self', 'vtt' and other implicit parameters from other kind of
parameters.
Reviewers: rjmccall, aaron.ballman
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D33735
llvm-svn: 305075
Amongst other, this will help LTO to correctly handle/honor files
compiled with O0, helping debugging failures.
It also seems in line with how we handle other options, like how
-fnoinline adds the appropriate attribute as well.
Differential Revision: https://reviews.llvm.org/D28404
llvm-svn: 304127
This patch implements codegen for the reduction clause on
any teams construct for elementary data types. It builds
on parallel reductions on the GPU. Subsequently,
the team master writes to a unique location in a global
memory scratchpad. The last team to do so loads and
reduces this array to calculate the final result.
This patch emits two helper functions that are used by
the OpenMP runtime on the GPU to perform reductions across
teams.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29879
llvm-svn: 295335
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 is a simple patch to teach OpenMP codegen to emit the construct
in Generic mode.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29143
llvm-svn: 293183
This patch adds support for the proc_bind clause on the Spmd construct
'target parallel' on the NVPTX device. Since the parallel region is created
upon kernel launch, this clause can be safely ignored on the NVPTX device at
codegen time for level 0 parallelism.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29128
llvm-svn: 293069
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
This patch adds support for the Spmd construct 'target parallel' on the
NVPTX device. This involves ignoring the num_threads clause on the device
since the number of threads in this combined construct is already set on
the host through the call to __tgt_target_teams().
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29083
llvm-svn: 292999
This patch adds codegen for the 'target parallel' directive on the NVPTX
device. We term offload OpenMP directives such as 'target parallel' and
'target teams distribute parallel for' as SPMD constructs. SPMD constructs,
in contrast to Generic ones like the plain 'target', can never contain
a serial region.
SPMD constructs can be handled more efficiently on the GPU and do not
require the Warp Loop of the Generic codegen scheme. This patch adds
SPMD codegen support for 'target parallel' on the NVPTX device and can
be reused for other SPMD constructs.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28755
llvm-svn: 292428
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
Summary:
This patch introduces support for the execution of parallel constructs in a target
region on the NVPTX device. Parallel regions must be in the lexical scope of the
target directive.
The master thread in the master warp signals parallel work for worker threads in worker
warps on encountering a parallel region.
Note: The patch does not yet support capture of arguments in a parallel region so
the test cases are simple.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28145
llvm-svn: 291565
Alexey Bataev