gfx11: don't use LL for sendrecv (#853)

* gfx11: don't use LL for sendrecv

* Use builtin instead of inline asm

src/collectives/device/primitives.h

@@ -21,6 +21,7 @@
   } else { \
     const int w = threadIdx.x/WARP_SIZE; \
     const int wid = threadIdx.x%WARP_SIZE; \
+    __threadfence(); \
     if (wid == 0) { \
       barrier_next[w] += nthreads/WARP_SIZE; \
       atomicAdd((unsigned long long *)barriers, 1); \

src/collectives/device/prims_simple.h

@@ -159,17 +159,18 @@ private:
 
   template<int Recv, int Send>
   inline __device__ void postPeer(bool dataStored) {
+    if (Send && (flags & RolePostSend) && dataStored)
+#ifdef __GFX9__
+      __builtin_amdgcn_buffer_wbinvl1();
+#else
+      __threadfence_system();
+#endif
+
     if ((flags & Send*RolePostSend) && next_hdp_reg)
       STORE((unsigned int *)next_hdp_reg, 0x1);
 
     if (flags & (Recv*RolePostRecv | Send*RolePostSend)) {
       step += StepPerSlice;
-      if (Send && (flags & RolePostSend) && dataStored)
-#ifdef __GFX9__
-        __asm__ __volatile__("buffer_wbinvl1_vol");
-#else
-        __threadfence_system();
-#endif
       STORE(connStepPtr, step);
     }
   }