diff --git a/llvm/include/llvm/Transforms/Scalar/LowerAtomic.h b/llvm/include/llvm/Transforms/Scalar/LowerAtomic.h
--- a/llvm/include/llvm/Transforms/Scalar/LowerAtomic.h
+++ b/llvm/include/llvm/Transforms/Scalar/LowerAtomic.h
@@ -14,6 +14,7 @@
 #ifndef LLVM_TRANSFORMS_SCALAR_LOWERATOMIC_H
 #define LLVM_TRANSFORMS_SCALAR_LOWERATOMIC_H
 
+#include "llvm/IR/Instructions.h"
 #include "llvm/IR/PassManager.h"
 
 namespace llvm {
@@ -24,6 +25,11 @@
   PreservedAnalyses run(Function &F, FunctionAnalysisManager &);
   static bool isRequired() { return true; }
 };
+
+/// Convert the given RMWI into primitive load and stores,
+/// assuming that doing so is legal. Return true if the lowering
+/// succeeds.
+bool LowerAtomicRMWInst(AtomicRMWInst *RMWI);
 }
 
 #endif // LLVM_TRANSFORMS_SCALAR_LOWERATOMIC_H
diff --git a/llvm/lib/Target/NVPTX/CMakeLists.txt b/llvm/lib/Target/NVPTX/CMakeLists.txt
--- a/llvm/lib/Target/NVPTX/CMakeLists.txt
+++ b/llvm/lib/Target/NVPTX/CMakeLists.txt
@@ -12,6 +12,7 @@
 
 set(NVPTXCodeGen_sources
   NVPTXAllocaHoisting.cpp
+  NVPTXAtomicLower.cpp
   NVPTXAsmPrinter.cpp
   NVPTXAssignValidGlobalNames.cpp
   NVPTXFrameLowering.cpp
diff --git a/llvm/lib/Target/NVPTX/NVPTXAtomicLower.h b/llvm/lib/Target/NVPTX/NVPTXAtomicLower.h
new file mode 100644
--- /dev/null
+++ b/llvm/lib/Target/NVPTX/NVPTXAtomicLower.h
@@ -0,0 +1,22 @@
+//===-- NVPTXAtomicLower.h - Lower atomics of local memory ------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//  Lower atomics of local memory to simple load/stores
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_NVPTX_NVPTXATOMICLOWER_H
+#define LLVM_LIB_TARGET_NVPTX_NVPTXATOMICLOWER_H
+
+namespace llvm {
+class FunctionPass;
+
+extern FunctionPass *createNVPTXAtomicLowerPass();
+} // end namespace llvm
+
+#endif
diff --git a/llvm/lib/Target/NVPTX/NVPTXAtomicLower.cpp b/llvm/lib/Target/NVPTX/NVPTXAtomicLower.cpp
new file mode 100644
--- /dev/null
+++ b/llvm/lib/Target/NVPTX/NVPTXAtomicLower.cpp
@@ -0,0 +1,74 @@
+//===-- NVPTXAtomicLower.cpp - Lower atomics of local memory ----*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//  Lower atomics of local memory to simple load/stores
+//
+//===----------------------------------------------------------------------===//
+
+#include "NVPTXAtomicLower.h"
+#include "llvm/CodeGen/StackProtector.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/Transforms/Scalar/LowerAtomic.h"
+
+#include "MCTargetDesc/NVPTXBaseInfo.h"
+using namespace llvm;
+
+namespace {
+// Hoisting the alloca instructions in the non-entry blocks to the entry
+// block.
+class NVPTXAtomicLower : public FunctionPass {
+public:
+  static char ID; // Pass ID
+  NVPTXAtomicLower() : FunctionPass(ID) {}
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+  }
+
+  StringRef getPassName() const override {
+    return "NVPTX lower atomics of local memory";
+  }
+
+  bool runOnFunction(Function &F) override;
+};
+} // namespace
+
+bool NVPTXAtomicLower::runOnFunction(Function &F) {
+  SmallVector<AtomicRMWInst *, 3> LocalMemoryAtomics;
+  for (BasicBlock &BB : F) {
+    for (Instruction &I : BB) {
+      if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(&I)) {
+        if (RMWI->getPointerAddressSpace() == ADDRESS_SPACE_LOCAL) {
+          LocalMemoryAtomics.push_back(RMWI);
+        }
+      }
+    }
+  }
+  bool Changed = false;
+  for (AtomicRMWInst *RMWI : LocalMemoryAtomics) {
+    Changed |= LowerAtomicRMWInst(RMWI);
+  }
+  return Changed;
+}
+
+char NVPTXAtomicLower::ID = 0;
+
+namespace llvm {
+void initializeNVPTXAtomicLowerPass(PassRegistry &);
+}
+
+INITIALIZE_PASS(NVPTXAtomicLower, "nvptx-atomic-lower",
+                "Lower atomics of local memory to simple load/stores", false,
+                false)
+
+FunctionPass *llvm::createNVPTXAtomicLowerPass() {
+  return new NVPTXAtomicLower();
+}
diff --git a/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp b/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
--- a/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp
@@ -13,6 +13,7 @@
 #include "NVPTXTargetMachine.h"
 #include "NVPTX.h"
 #include "NVPTXAllocaHoisting.h"
+#include "NVPTXAtomicLower.h"
 #include "NVPTXLowerAggrCopies.h"
 #include "NVPTXTargetObjectFile.h"
 #include "NVPTXTargetTransformInfo.h"
@@ -65,6 +66,7 @@
 void initializeNVVMReflectPass(PassRegistry&);
 void initializeGenericToNVVMPass(PassRegistry&);
 void initializeNVPTXAllocaHoistingPass(PassRegistry &);
+void initializeNVPTXAtomicLowerPass(PassRegistry &);
 void initializeNVPTXAssignValidGlobalNamesPass(PassRegistry&);
 void initializeNVPTXLowerAggrCopiesPass(PassRegistry &);
 void initializeNVPTXLowerArgsPass(PassRegistry &);
@@ -86,6 +88,7 @@
   initializeGenericToNVVMPass(PR);
   initializeNVPTXAllocaHoistingPass(PR);
   initializeNVPTXAssignValidGlobalNamesPass(PR);
+  initializeNVPTXAtomicLowerPass(PR);
   initializeNVPTXLowerArgsPass(PR);
   initializeNVPTXLowerAllocaPass(PR);
   initializeNVPTXLowerAggrCopiesPass(PR);
@@ -252,6 +255,7 @@
   addPass(createSROAPass());
   addPass(createNVPTXLowerAllocaPass());
   addPass(createInferAddressSpacesPass());
+  addPass(createNVPTXAtomicLowerPass());
 }
 
 void NVPTXPassConfig::addStraightLineScalarOptimizationPasses() {
diff --git a/llvm/lib/Transforms/Scalar/LowerAtomic.cpp b/llvm/lib/Transforms/Scalar/LowerAtomic.cpp
--- a/llvm/lib/Transforms/Scalar/LowerAtomic.cpp
+++ b/llvm/lib/Transforms/Scalar/LowerAtomic.cpp
@@ -40,7 +40,7 @@
   return true;
 }
 
-static bool LowerAtomicRMWInst(AtomicRMWInst *RMWI) {
+bool llvm::LowerAtomicRMWInst(AtomicRMWInst *RMWI) {
   IRBuilder<> Builder(RMWI);
   Value *Ptr = RMWI->getPointerOperand();
   Value *Val = RMWI->getValOperand();
diff --git a/llvm/test/CodeGen/NVPTX/atomic-lower.ll b/llvm/test/CodeGen/NVPTX/atomic-lower.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/atomic-lower.ll
@@ -0,0 +1,14 @@
+; RUN: opt < %s -S -nvptx-atomic-lower | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
+target triple = "nvptx64-unknown-unknown"
+
+define double @kernel(double addrspace(5)* %ptr, double %val) {
+  %res = atomicrmw fadd double addrspace(5)* %ptr, double %val monotonic, align 8
+  ret double %res
+; CHECK:   %1 = load double, double addrspace(5)* %ptr, align 8
+; CHECK-NEXT:   %2 = fadd double %1, %val
+; CHECK-NEXT:   store double %2, double addrspace(5)* %ptr, align 8
+; CHECK-NEXT:   ret double %1
+}
+