Merge pull request #34386 from VinInn/fenceforampere

enforce memory order to avoid crash on Ampere

Author: cmsbuild

RecoPixelVertexing/PixelTriplets/plugins/GPUCACell.h

@@ -69,6 +69,7 @@ class GPUCACell {
       auto i = cellNeighbors.extend();  // maybe wasted....
       if (i > 0) {
         cellNeighbors[i].reset();
+        __threadfence();
 #ifdef __CUDACC__
         auto zero = (PtrAsInt)(&cellNeighbors[0]);
         atomicCAS((PtrAsInt*)(&theOuterNeighbors),
@@ -89,6 +90,7 @@ class GPUCACell {
       auto i = cellTracks.extend();  // maybe wasted....
       if (i > 0) {
         cellTracks[i].reset();
+        __threadfence();
 #ifdef __CUDACC__
         auto zero = (PtrAsInt)(&cellTracks[0]);
         atomicCAS((PtrAsInt*)(&theTracks), zero, (PtrAsInt)(&cellTracks[i]));  // if fails we cannot give "i" back...