[OpenMP] Parallel reduction on the NVPTX device.

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

1 files changed, 26 insertions, 7 deletions

diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.h b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
index d1ff0e8a24a..ae25e94759e 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
@@ -67,12 +67,6 @@ private:
   /// \brief Signal termination of Spmd mode execution.
   void emitSpmdEntryFooter(CodeGenFunction &CGF, EntryFunctionState &EST);
 
-  /// \brief Returns specified OpenMP runtime function for the current OpenMP
-  /// implementation.  Specialized for the NVPTX device.
-  /// \param Function OpenMP runtime function.
-  /// \return Specified function.
-  llvm::Constant *createNVPTXRuntimeFunction(unsigned Function);
-
   //
   // Base class overrides.
   //
@@ -248,7 +242,32 @@ public:
                         ArrayRef<llvm::Value *> CapturedVars,
                         const Expr *IfCond) override;
 
-public:
+  /// Emit a code for reduction clause.
+  ///
+  /// \param Privates List of private copies for original reduction arguments.
+  /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
+  /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
+  /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
+  /// or 'operator binop(LHS, RHS)'.
+  /// \param Options List of options for reduction codegen:
+  ///     WithNowait true if parent directive has also nowait clause, false
+  ///     otherwise.
+  ///     SimpleReduction Emit reduction operation only. Used for omp simd
+  ///     directive on the host.
+  ///     ReductionKind The kind of reduction to perform.
+  virtual void emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
+                             ArrayRef<const Expr *> Privates,
+                             ArrayRef<const Expr *> LHSExprs,
+                             ArrayRef<const Expr *> RHSExprs,
+                             ArrayRef<const Expr *> ReductionOps,
+                             ReductionOptionsTy Options) override;
+
+  /// Returns specified OpenMP runtime function for the current OpenMP
+  /// implementation.  Specialized for the NVPTX device.
+  /// \param Function OpenMP runtime function.
+  /// \return Specified function.
+  llvm::Constant *createNVPTXRuntimeFunction(unsigned Function);
+
   /// Target codegen is specialized based on two programming models: the
   /// 'generic' fork-join model of OpenMP, and a more GPU efficient 'spmd'
   /// model for constructs like 'target parallel' that support it.