Initial resize scheduler

CMakeLists.txt

@@ -231,8 +231,10 @@ list(APPEND NVFUSER_SRCS
   ${NVFUSER_SRCS_DIR}/scheduler/reduction_utils.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/registry.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/registry_utils.cpp
+  ${NVFUSER_SRCS_DIR}/scheduler/resize.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/runtime_info.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/scheduler_types.cpp
+  ${NVFUSER_SRCS_DIR}/scheduler/tools/domain_map.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/tools/inlining.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/tools/loop_domain_scheduler.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/tools/maxinfo_propagator.cpp

csrc/device_lower/pass/expr_sort.cpp

@@ -1088,7 +1088,7 @@ bool ExprSegmentationSorter::interIterUpdate() {
     NVF_ERROR(
         !fallback_mode_enabled_,
         "Couldn't succcessfully sort out the fusion expressions. ",
-        "There are remaining connections of the heirarchical segmentation which should have been ",
+        "There are remaining connections of the hierarchical segmentation which should have been ",
         "flattened to a single ordered group, or disjoint ordered groups.\n",
         toString());
     // We didn't finish, but we haven't tried the fallback, try again with that.

csrc/scheduler/compile_time_info.h

@@ -10,6 +10,7 @@
 #include <fusion.h>
 #include <scheduler/pointwise_utils.h>
 #include <scheduler/scheduler_types.h>
+#include <scheduler/tools/domain_map.h>
 #include <scheduler/utils.h>
 #include <scheduler/vectorize_helper.h>
 
@@ -54,7 +55,7 @@ enum class CompileTimeEntryType {
 //!  stores the domain map of a fusion.
 class DomainMap {
  public:
-  using DataType = pointwise_utils::DomainMap;
+  using DataType = scheduler_tools::DomainMap;
   static const CompileTimeEntryType EntryType =
       CompileTimeEntryType::DOMAIN_MAP;
 };
@@ -63,7 +64,7 @@ class DomainMap {
 //!  stores the domain map of a fusion, used by transpose scheduler.
 class TransposeDomainMap {
  public:
-  using DataType = pointwise_utils::DomainMap;
+  using DataType = scheduler_tools::DomainMap;
   static const CompileTimeEntryType EntryType =
       CompileTimeEntryType::TRANSPOSE_DOMAIN_MAP;
 };

csrc/scheduler/pointwise.cpp

@@ -29,37 +29,6 @@ namespace {
 // Unused at the moment, commenting for clang tidy
 constexpr int64_t kThreadX = 128;
 
-class DomainMap : public pointwise_utils::DomainMap {
- public:
-  using pointwise_utils::DomainMap::DomainMap;
-
-  // The pointwise scheduler heuristics requires a minimum number of axes.
-  // The output reference tensor should respect this requirement.
-  TensorView* findReferenceTensorView(int64_t minimum_num_axes = 0) const {
-    TensorView* result = nullptr;
-    int64_t max_dims = -1;
-    for (auto output_tv :
-         ir_utils::filterByType<TensorView>(fusion_->outputs())) {
-      if (isValidReference(output_tv) &&
-          hasMinimumSize(output_tv, minimum_num_axes) &&
-          !output_tv->isFusionInput()) {
-        int64_t n_dims = pointwise_utils::nRootDims(output_tv);
-        if (n_dims > max_dims) {
-          result = output_tv;
-          max_dims = n_dims;
-        }
-      }
-    }
-    return result;
-  }
-
- private:
-  bool hasMinimumSize(TensorView* tv, int64_t num_axes) const {
-    NVF_ERROR(tv != nullptr);
-    return (num_axes == 0 || (int64_t)tv->getLogicalDomain().size() > num_axes);
-  }
-};
-
 } // namespace
 
 std::unique_ptr<PointwiseParams> getPointwiseHeuristics(
@@ -79,14 +48,16 @@ std::unique_ptr<PointwiseParams> getPointwiseHeuristics(
 
   auto domain_map_entry =
       HeuristicDataCacheEntry<HeuristicCompileTime::DomainMap>(
-          data_cache,
-          [fusion]() { return std::make_unique<DomainMap>(fusion); });
-  const auto& domain_map = dynamic_cast<DomainMap&>(domain_map_entry.get());
+          data_cache, [fusion]() {
+            return std::make_unique<pointwise_utils::DomainMap>(fusion);
+          });
+  const auto& domain_map =
+      dynamic_cast<pointwise_utils::DomainMap&>(domain_map_entry.get());
 
   auto largest_out_entry =
       HeuristicDataCacheEntry<HeuristicCompileTime::ReferenceTensors>(
           data_cache, [&domain_map]() {
-            std::vector<TensorView*> data{domain_map.findReferenceTensorView()};
+            std::vector<TensorView*> data{domain_map.findReferenceTensor()};
             return std::make_unique<std::vector<TensorView*>>(std::move(data));
           });
   TensorView* largest_out = largest_out_entry.get()[0];
@@ -432,19 +403,11 @@ std::unique_ptr<PointwiseParams> getPointwiseHeuristics(
   return params;
 }
 
-// Return reference tensor view.
-TensorView* getReferenceTensorView(Fusion* fusion) {
-  FusionGuard fg(fusion);
-  DomainMap domain_map(fusion);
-  auto reference_tv = domain_map.findReferenceTensorView();
-  return reference_tv;
-}
-
 //! Utility for canSchedule interface to check if this fusion has
 //!  a fully broadcasted reference tensor, which is necessary for
 //!  the pointwise scheduler.
 bool hasReferenceTensorView(Fusion* fusion) {
-  return getReferenceTensorView(fusion) != nullptr;
+  return pointwise_utils::getReferenceTensor(fusion) != nullptr;
 }
 
 bool PointWiseScheduler::canScheduleCompileTime(Fusion* fusion) {
@@ -541,7 +504,7 @@ void schedulePointwise(Fusion* fusion, const PointwiseParams* pparams) {
     return;
   }
 
-  TensorView* reference_tv = getReferenceTensorView(fusion);
+  TensorView* reference_tv = pointwise_utils::getReferenceTensor(fusion);
 
   NVF_ERROR(
       reference_tv != nullptr,