Improve barrier implementation

[ROCm/rccl commit: adb8d63352]

projects/rccl/src/include/devcomm.h

@@ -18,8 +18,8 @@
 
 // Convert volatile access to atomic
 #if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
-#define LOAD(VAR) __atomic_load_n((VAR), __ATOMIC_SEQ_CST)
-#define STORE(DST, SRC) __atomic_store_n((DST), (SRC), __ATOMIC_SEQ_CST)
+#define LOAD(VAR) __atomic_load_n((VAR), __ATOMIC_ACQUIRE)
+#define STORE(DST, SRC) __atomic_store_n((DST), (SRC), __ATOMIC_RELEASE)
 #else
 #define LOAD(VAR) *(VAR)
 #define STORE(DST, SRC) *(DST) = (SRC)