[OPENMP][NVPTX]Improved interwarp copy function.
Inlined runtime with the current implementation of the interwarp copy
function leads to the undefined behavior because of the not quite
correct implementation of the barriers. Start using generic
__kmpc_barier function instead of the custom made barriers.
diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index a7b3c16..b055132 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -189,13 +189,6 @@
SharedMemorySize = 128,
};
-enum NamedBarrier : unsigned {
- /// Synchronize on this barrier #ID using a named barrier primitive.
- /// Only the subset of active threads in a parallel region arrive at the
- /// barrier.
- NB_Parallel = 1,
-};
-
static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
RefExpr = RefExpr->IgnoreParens();
if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
@@ -655,26 +648,9 @@
CGF.EmitRuntimeCall(F);
}
-/// Get barrier #ID to synchronize selected (multiple of warp size) threads in
-/// a CTA.
-static void getNVPTXBarrier(CodeGenFunction &CGF, int ID,
- llvm::Value *NumThreads) {
- CGBuilderTy &Bld = CGF.Builder;
- llvm::Value *Args[] = {Bld.getInt32(ID), NumThreads};
- llvm::Function *F = llvm::Intrinsic::getDeclaration(
- &CGF.CGM.getModule(), llvm::Intrinsic::nvvm_barrier);
- F->addFnAttr(llvm::Attribute::Convergent);
- CGF.EmitRuntimeCall(F, Args);
-}
-
/// Synchronize all GPU threads in a block.
static void syncCTAThreads(CodeGenFunction &CGF) { getNVPTXCTABarrier(CGF); }
-/// Synchronize worker threads in a parallel region.
-static void syncParallelThreads(CodeGenFunction &CGF, llvm::Value *NumThreads) {
- return getNVPTXBarrier(CGF, NB_Parallel, NumThreads);
-}
-
/// Get the value of the thread_limit clause in the teams directive.
/// For the 'generic' execution mode, the runtime encodes thread_limit in
/// the launch parameters, always starting thread_limit+warpSize threads per
@@ -3272,14 +3248,10 @@
CGF.EmitBlock(MergeBB);
- Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
- llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
- AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
-
- llvm::Value *NumActiveThreads = Bld.CreateNSWMul(
- NumWarpsVal, getNVPTXWarpSize(CGF), "num_active_threads");
- // named_barrier_sync(ParallelBarrierID, num_active_threads)
- syncParallelThreads(CGF, NumActiveThreads);
+ // kmpc_barrier.
+ CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
+ /*EmitChecks=*/false,
+ /*ForceSimpleCall=*/true);
//
// Warp 0 copies reduce element from transfer medium.
@@ -3288,6 +3260,10 @@
llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");
+ Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
+ llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
+ AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
+
// Up to 32 threads in warp 0 are active.
llvm::Value *IsActiveThread =
Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
@@ -3329,7 +3305,10 @@
// While warp 0 copies values from transfer medium, all other warps must
// wait.
- syncParallelThreads(CGF, NumActiveThreads);
+ // kmpc_barrier.
+ CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
+ /*EmitChecks=*/false,
+ /*ForceSimpleCall=*/true);
if (NumIters > 1) {
Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
diff --git a/clang/test/OpenMP/nvptx_target_parallel_reduction_codegen.cpp b/clang/test/OpenMP/nvptx_target_parallel_reduction_codegen.cpp
index 965ff9e..dd93b0c 100644
--- a/clang/test/OpenMP/nvptx_target_parallel_reduction_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_target_parallel_reduction_codegen.cpp
@@ -209,9 +209,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -231,7 +230,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: [[NEXT:%.+]] = add nsw i32 [[CNT]], 1
// CHECK: store i32 [[NEXT]], i32* [[CNT_ADDR]],
// CHECK: br label
@@ -446,9 +445,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -467,7 +465,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: [[IS_WARP_MASTER:%.+]] = icmp eq i32 [[LANEID]], 0
// CHECK: br i1 [[IS_WARP_MASTER]], label {{%?}}[[DO_COPY:.+]], label {{%?}}[[COPY_ELSE:.+]]
//
@@ -486,9 +484,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -507,7 +504,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: ret
@@ -761,9 +758,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -782,7 +778,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: [[IS_WARP_MASTER:%.+]] = icmp eq i32 [[LANEID]], 0
// CHECK: br i1 [[IS_WARP_MASTER]], label {{%?}}[[DO_COPY:.+]], label {{%?}}[[COPY_ELSE:.+]]
//
@@ -802,9 +798,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -824,7 +819,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: ret
#endif
diff --git a/clang/test/OpenMP/nvptx_teams_reduction_codegen.cpp b/clang/test/OpenMP/nvptx_teams_reduction_codegen.cpp
index 97a1288..65c147b 100644
--- a/clang/test/OpenMP/nvptx_teams_reduction_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_teams_reduction_codegen.cpp
@@ -412,9 +412,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -433,7 +432,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: [[IS_WARP_MASTER:%.+]] = icmp eq i32 [[LANEID]], 0
// CHECK: br i1 [[IS_WARP_MASTER]], label {{%?}}[[DO_COPY:.+]], label {{%?}}[[COPY_ELSE:.+]]
//
@@ -453,9 +452,8 @@
//
// Barrier after copy to shared memory storage medium.
// CHECK: [[COPY_CONT]]
- // CHECK: [[WS:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize()
- // CHECK: [[ACTIVE_THREADS:%.+]] = mul nsw i32 [[ACTIVE_WARPS:%.+]], [[WS]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
+ // CHECK: [[ACTIVE_WARPS:%.+]] = load i32, i32*
//
// Read into warp 0.
// CHECK: [[IS_W0_ACTIVE_THREAD:%.+]] = icmp ult i32 [[TID:%.+]], [[ACTIVE_WARPS]]
@@ -475,7 +473,7 @@
// CHECK: br label {{%?}}[[READ_CONT]]
//
// CHECK: [[READ_CONT]]
- // CHECK: call void @llvm.nvvm.barrier(i32 1, i32 [[ACTIVE_THREADS]])
+ // CHECK: call void @__kmpc_barrier(%struct.ident_t* @
// CHECK: ret
#endif
