[AMD][BACKEND] Bugfix to small tile pingpong (#5759)

Do not simply append dot op next to the memory ops.
It can reorder the dependency to the dot op.
Fix test failure in test_matmul.

Author: jungpark-mlir

test/TritonGPU/amd/amd-block-pingpong.mlir

@@ -4,11 +4,11 @@
 //CHECK: ttg.local_load
 //CHECK: rocdl.s.setprio 1
 //CHECK: tt.load
-//CHECK: rocdl.sched.barrier 0
+//CHECK: rocdl.sched.barrier
 //CHECK: ttg.local_load
 //CHECK: rocdl.s.setprio 0
 //CHECK: tt.load
-//CHECK: rocdl.sched.barrier 0
+//CHECK: rocdl.sched.barrier
 //CHECK: rocdl.s.setprio 1
 //CHECK: tt.dot
 //CHECK: rocdl.s.setprio 0

third_party/amd/lib/TritonAMDGPUTransforms/BlockPingpong.cpp

@@ -119,6 +119,13 @@ void Pingponger::appendOpWithPrio(OpBuilder &builder, Operation *op,
 // high-level operations, inserting `setPrio` also has a same effect of
 // instruction scheduling boundary, too.
 void Pingponger::transformOnePPClusters(OpBuilder &builder, Location loc) {
+  auto dotLoc = dotOps[0]->getPrevNode();
+  // sched barrier to prevent memory ops from cross but leave other ops to be
+  // scheduled across the barrier.
+  auto preDotBar = builder.create<ROCDL::SchedBarrier>(loc, 1);
+  updateOpInsertion(dotLoc);
+  appendOp(preDotBar);
+
   // Memory cluster #0
   updateOpInsertion(lLoadOps[0]);
   appendOp(builder.create<ROCDL::SetPrioOp>(loc, highPriority));
@@ -127,9 +134,9 @@ void Pingponger::transformOnePPClusters(OpBuilder &builder, Location loc) {
   appendOp(lLoadOps[1]);
   appendOp(builder.create<ROCDL::SetPrioOp>(loc, lowPriority));
   appendOp(gLoadOps[1]);
-  appendOp(builder.create<ROCDL::SchedBarrier>(loc, 0));
 
   // Dot cluster #0
+  updateOpInsertion(preDotBar);
   appendOpWithPrio(builder, dotOps[0], loc);
 }