WIP

csrc/scheduler/resize.cpp

@@ -137,11 +137,11 @@ void ResizeScheduler::schedule(Fusion* fusion, const HeuristicParams* params) {
   FUSER_PERF_SCOPE("ResizeScheduler::schedule");
 
   FusionGuard fg(fusion);
-
+#if 0
   fusion->printMath();
   fusion->print();
   std::cout << std::endl;
-
+#endif
   scheduler_utils::clearMemorySpace(fusion);
 
   scheduler_utils::cacheInputs(fusion, true);
@@ -153,7 +153,7 @@ void ResizeScheduler::schedule(Fusion* fusion, const HeuristicParams* params) {
 
     scheduler_tools::propagateResizeToInputs(expr);
   }
-
+#if 0
   fusion->print();
   std::cout << std::endl;
 
@@ -173,14 +173,28 @@ void ResizeScheduler::schedule(Fusion* fusion, const HeuristicParams* params) {
       std::cerr << expr->toString(4);
     }
   }
-
+#endif
   auto ref_tv = getReferenceTensor(fusion);
 
   // Just simple scheduling for now.
   // TODO: Do something smarter. Can just use the pointwise scheduler?
-  ref_tv->flatten();
-  ref_tv->split(0, 128);
-  ref_tv->split(0, 1 << 14);
+
+  // Make sure the DID ID located at the outermost position
+  int64_t reorder_pos = 0;
+  std::unordered_map<int64_t, int64_t> old2new;
+  for (const auto i : c10::irange(ref_tv->getLoopDomain().size())) {
+    if (isParallelTypeDeviceDim(ref_tv->axis((int64_t)i)->getParallelType())) {
+      old2new.emplace((int64_t)i, reorder_pos);
+      ++reorder_pos;
+    }
+  }
+  ref_tv->reorder(old2new);
+
+  // Schedule only the remaining IDs
+  const auto outermost_pos = (int64_t)old2new.size();
+  ref_tv->flatten(outermost_pos);
+  ref_tv->split(outermost_pos, 128);
+  ref_tv->split(outermost_pos, 1 << 14);
   ref_tv->axis(-1)->parallelize(ParallelType::TIDx);
   ref_tv->axis(-2)->parallelize(ParallelType::BIDx);