[Openmp-commits] [openmp] 219355d - [Libomptarget] Use scoped atomics in the device runtime (#75834)

[via Openmp-commits]

Author: Joseph Huber
Date: 2023-12-19T14:30:34-06:00
New Revision: 219355d4c0d2b6e2c0d5e022f8b7a78c1e9ce53f

URL: https://github.com/llvm/llvm-project/commit/219355d4c0d2b6e2c0d5e022f8b7a78c1e9ce53f
DIFF: https://github.com/llvm/llvm-project/commit/219355d4c0d2b6e2c0d5e022f8b7a78c1e9ce53f.diff

LOG: [Libomptarget] Use scoped atomics in the device runtime (#75834)

Summary:
A recent patch allowed us to easily replace GNU atomics with scoped
variants that make use of the backend's handling for more permissive
scopes. The default is full "system" scope, that means the atomic
operation must be consistent with operations that may happen on the
host's memory. This is generally only required for processes that are
communicating with something via global fine-grained memory. This patch
uses these atomics to make everything device scoped, as nothing in the
OpenMP runtime should depend on the host.

This is only provided as a very new clang extension but the DeviceRTL is
only compiled with clang so it is always available.

Added: 
    

Modified: 
    openmp/libomptarget/DeviceRTL/src/Synchronization.cpp

Removed: 
    


################################################################################
diff  --git a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
index ad3b1cad419436..57f98a353589c4 100644
--- a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
@@ -34,7 +34,8 @@ uint32_t atomicInc(uint32_t *Address, uint32_t Val, atomic::OrderingTy Ordering,
 
 template <typename Ty>
 Ty atomicAdd(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_add(Address, Val, Ordering);
+  return __scoped_atomic_fetch_add(Address, Val, Ordering,
+                                   __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
@@ -56,25 +57,28 @@ template <typename Ty> Ty atomicLoad(Ty *Address, atomic::OrderingTy Ordering) {
 
 template <typename Ty>
 void atomicStore(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  __atomic_store_n(Address, Val, Ordering);
+  __scoped_atomic_store_n(Address, Val, Ordering, __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
 bool atomicCAS(Ty *Address, Ty ExpectedV, Ty DesiredV,
                atomic::OrderingTy OrderingSucc,
                atomic::OrderingTy OrderingFail) {
-  return __atomic_compare_exchange(Address, &ExpectedV, &DesiredV, false,
-                                   OrderingSucc, OrderingFail);
+  return __scoped_atomic_compare_exchange(Address, &ExpectedV, &DesiredV, false,
+                                          OrderingSucc, OrderingFail,
+                                          __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
 Ty atomicMin(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_min(Address, Val, Ordering);
+  return __scoped_atomic_fetch_min(Address, Val, Ordering,
+                                   __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
 Ty atomicMax(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_max(Address, Val, Ordering);
+  return __scoped_atomic_fetch_max(Address, Val, Ordering,
+                                   __MEMORY_SCOPE_DEVICE);
 }
 
 // TODO: Implement this with __atomic_fetch_max and remove the duplication.
@@ -94,23 +98,26 @@ Ty atomicMaxFP(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
 
 template <typename Ty>
 Ty atomicOr(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_or(Address, Val, Ordering);
+  return __scoped_atomic_fetch_or(Address, Val, Ordering,
+                                  __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
 Ty atomicAnd(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_and(Address, Val, Ordering);
+  return __scoped_atomic_fetch_and(Address, Val, Ordering,
+                                   __MEMORY_SCOPE_DEVICE);
 }
 
 template <typename Ty>
 Ty atomicXOr(Ty *Address, Ty Val, atomic::OrderingTy Ordering) {
-  return __atomic_fetch_xor(Address, Val, Ordering);
+  return __scoped_atomic_fetch_xor(Address, Val, Ordering,
+                                   __MEMORY_SCOPE_DEVICE);
 }
 
 uint32_t atomicExchange(uint32_t *Address, uint32_t Val,
                         atomic::OrderingTy Ordering) {
   uint32_t R;
-  __atomic_exchange(Address, &Val, &R, Ordering);
+  __scoped_atomic_exchange(Address, &Val, &R, Ordering, __MEMORY_SCOPE_DEVICE);
   return R;
 }
 ///}