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
Summary:
This patch adds two fields to the offload entry descriptor. One field is meant to signal Ctors/Dtors and `link` global variables, and the other is reserved for runtime library use.
Currently, these fields are only filled with zeros in the current code generation, but that will change when `declare target` is added.
The reason, we are adding these fields now is to make the code generation consistent with the runtime library proposal under review in https://reviews.llvm.org/D14031.
Reviewers: ABataev, hfinkel, carlo.bertolli, kkwli0, arpith-jacob, Hahnfeld
Subscribers: cfe-commits, caomhin, jholewinski
Differential Revision: https://reviews.llvm.org/D28298
llvm-svn: 291124
This patch includes updates for codegen of the target region for the NVPTX
device. It moves initializers from the compiler to the runtime and updates
the worker loop to assume parallel work is retrieved from the runtime. A
subsequent patch will update the codegen to retrieve the parallel work using
calls to the runtime. It includes the removal of the inline attribute
for the worker loop and disabling debug info in it.
This allows codegen for a target directive and serial execution on the
NVPTX device.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28125
llvm-svn: 291121
This patch includes updates for codegen of the target region for the NVPTX
device. It moves initializers from the compiler to the runtime and updates
the worker loop to assume parallel work is retrieved from the runtime. A
subsequent patch will update the codegen to retrieve the parallel work using
calls to the runtime. It includes the removal of the inline attribute
for the worker loop and disabling debug info in it.
This allows codegen for a target directive and serial execution on the
NVPTX device.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28125
llvm-svn: 290983
This patch cleans up private methods for NVPTX OpenMP codegen. It converts private
members to static functions to follow the coding style of CGOpenMPRuntime.cpp and
declutter the header file.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28124
llvm-svn: 290904
-fno-inline-functions, -O0, and optnone.
These were really, really tangled together:
- We used the noinline LLVM attribute for -fno-inline
- But not for -fno-inline-functions (breaking LTO)
- But we did use it for -finline-hint-functions (yay, LTO is happy!)
- But we didn't for -O0 (LTO is sad yet again...)
- We had weird structuring of CodeGenOpts with both an inlining
enumeration and a boolean. They interacted in weird ways and
needlessly.
- A *lot* of set smashing went on with setting these, and then got worse
when we considered optnone and other inlining-effecting attributes.
- A bunch of inline affecting attributes were managed in a completely
different place from -fno-inline.
- Even with -fno-inline we failed to put the LLVM noinline attribute
onto many generated function definitions because they didn't show up
as AST-level functions.
- If you passed -O0 but -finline-functions we would run the normal
inliner pass in LLVM despite it being in the O0 pipeline, which really
doesn't make much sense.
- Lastly, we used things like '-fno-inline' to manipulate the pass
pipeline which forced the pass pipeline to be much more
parameterizable than it really needs to be. Instead we can *just* use
the optimization level to select a pipeline and control the rest via
attributes.
Sadly, this causes a bunch of churn in tests because we don't run the
optimizer in the tests and check the contents of attribute sets. It
would be awesome if attribute sets were a bit more FileCheck friendly,
but oh well.
I think this is a significant improvement and should remove the semantic
need to change what inliner pass we run in order to comply with the
requested inlining semantics by relying completely on attributes. It
also cleans up tho optnone and related handling a bit.
One unfortunate aspect of this is that for generating alwaysinline
routines like those in OpenMP we end up removing noinline and then
adding alwaysinline. I tried a bunch of other approaches, but because we
recompute function attributes from scratch and don't have a declaration
here I couldn't find anything substantially cleaner than this.
Differential Revision: https://reviews.llvm.org/D28053
llvm-svn: 290398
program
Offload related code is not quite ready yet, but some simple examples
must not crash the compiler. Patch fixes the problem in offloading code
with exceptions.
llvm-svn: 290364
This patch implements the teams directive for the NVPTX backend. It is different from the host code generation path as it:
Does not call kmpc_fork_teams. All necessary teams and threads are started upon touching the target region, when launching a CUDA kernel, and their execution is coordinated through sequential and parallel regions within the target region.
Does not call kmpc_push_num_teams even if a num_teams of thread_limit clause is present. Setting the number of teams and the thread limit is implemented by the nvptx-related runtime.
Please note that I am now passing a Clang Expr * to emitPushNumTeams instead of the originally chosen llvm::Value * type. The reason for that is that I want to avoid emitting expressions for num_teams and thread_limit if they are not needed in the target region.
http://reviews.llvm.org/D17963
llvm-svn: 265304
Solution unifies interface of RegionCodeGenTy type to allow insert
runtime-specific code before/after main codegen action defined in
CGStmtOpenMP.cpp file. Runtime should not define its own RegionCodeGenTy
for general OpenMP directives, but must be allowed to insert its own
(required) code to support target specific codegen.
llvm-svn: 264700
Solution unifies interface of RegionCodeGenTy type to allow insert
runtime-specific code before/after main codegen action defined in
CGStmtOpenMP.cpp file. Runtime should not define its own RegionCodeGenTy
for general OpenMP directives, but must be allowed to insert its own
(required) code to support target specific codegen.
llvm-svn: 264576
Solution unifies interface of RegionCodeGenTy type to allow insert
runtime-specific code before/after main codegen action defined in
CGStmtOpenMP.cpp file. Runtime should not define its own RegionCodeGenTy
for general OpenMP directives, but must be allowed to insert its own
(required) code to support target specific codegen.
llvm-svn: 264569
Summary:
This patch adds base support for codegen of the target directive on the NVPTX device.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D17877
Reworked test case after buildbot failure on windows.
Updated patch to integrate r263837 and test case nvptx_target_firstprivate_codegen.cpp.
llvm-svn: 264018
Summary:
Reworked test case after buildbot failure on windows.
This patch adds base support for codegen of the target directive on the NVPTX device.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D17877
llvm-svn: 263783
Summary:
This patch adds base support for codegen of the target directive on the NVPTX device.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D17877
llvm-svn: 263587
Summary:
This patch adds base support for codegen of the target directive on the NVPTX device.
Reviewers: ABataev
Differential Revision: http://reviews.llvm.org/D17877
llvm-svn: 263552
Summary:
Different devices may in some cases require different code generation schemes in order to implement OpenMP. This is required not only for performance reasons, but also because it may not be possible to have the current (default) implementation working for these devices. E.g. GPU's cannot implement the same scheme a target such as powerpc or x86b would use, in the sense that it does not have the ability to fork threads, instead all the threads are always executing and need to be managed by the implementation.
This patch proposes a reorganization of the code in the OpenMP code generation to pave the way to have specialized implementation of OpenMP support. More than a "real" patch this is more a request for comments in order to understand if what is proposed is acceptable or if there are better/easier ways to do it.
In this patch part of the common OpenMP codegen infrastructure is moved to a new file under a new namespace (CGOpenMPCommon) so it can be shared between the default implementation and the specialized one. When CGOpenMPRuntime is created, an attempt to select a specialized implementation is done.
In the patch a specialization for nvptx targets is done which currently checks if the target is an OpenMP device and trap if it is not.
Let me know comments suggestions you may have.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev
Subscribers: Hahnfeld, cfe-commits, fraggamuffin, caomhin, jholewinski
Differential Revision: http://reviews.llvm.org/D16784
llvm-svn: 259977
Arpith Chacko Jacob