Limit max channels for ring graph on single node Rome (#347)

* Limit max channels for ring graph on single node Rome
* Partially revert "Use non-temporal access for streaming data (#341)"

tools/rccl-prim-test/copy_kernel.h

@@ -86,17 +86,16 @@ struct MULTI128 {
 
 inline __device__ void Fetch128(Pack128& v, const Pack128* p) {
 #if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
-  v.x = __builtin_nontemporal_load(&p->x);
-  v.y = __builtin_nontemporal_load(&p->y);
+  v.x = p->x;
+  v.y = p->y;
 #else
   asm volatile("ld.volatile.global.v2.u64 {%0,%1}, [%2];" : "=l"(v.x), "=l"(v.y) : "l"(p) : "memory");
 #endif
 }
-
 inline __device__ void Store128(Pack128* p, Pack128& v) {
 #if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
-  __builtin_nontemporal_store(v.x, &p->x);
-  __builtin_nontemporal_store(v.y, &p->y);
+  p->x = v.x;
+  p->y = v.y;
 #else
   asm volatile("st.volatile.global.v2.u64 [%0], {%1,%2};" :: "l"(p), "l"(v.x), "l"(v.y) : "memory");
 #endif

tools/rccl-prim-test/rccl_prim_test.cpp

@@ -29,6 +29,7 @@ THE SOFTWARE.
 #include <iostream> //cerr
 #include <unistd.h> //usleep
 #include <cstring>
+#include <hip/hip_runtime_api.h>
 #include <hip/hip_runtime.h>
 #include "copy_kernel.h"