Indexing refactor stage 1: remove reference tensor creation in all te…

…nsor indexing logic (csarofeen#1690)

Co-authored-by: Naoya Maruyama <naoyam@users.noreply.github.com>
Co-authored-by: Christian Sarofeen <csarofeen@nvidia.com>

build_variables.bzl

@@ -656,6 +656,7 @@ libtorch_cuda_core_sources = [
     "torch/csrc/jit/codegen/cuda/graph_fuser.cpp",
     "torch/csrc/jit/codegen/cuda/grouped_reduction.cpp",
     "torch/csrc/jit/codegen/cuda/index_compute.cpp",
+    "torch/csrc/jit/codegen/cuda/lower_index_compute.cpp",
     "torch/csrc/jit/codegen/cuda/index_reference_replay.cpp",
     "torch/csrc/jit/codegen/cuda/instrumentation.cpp",
     "torch/csrc/jit/codegen/cuda/ir_base_nodes.cpp",

torch/csrc/jit/codegen/cuda/compute_at_map.cpp

@@ -226,7 +226,8 @@ void IterDomainGraph::initializeId(
   }
 }
 
-ComputeAtMap::ComputeAtMap(Fusion* fusion) : id_graph_(fusion) {
+ComputeAtMap::ComputeAtMap(Fusion* fusion)
+    : id_graph_(fusion), fusion_(fusion) {
   build(fusion);
 }
 
@@ -257,6 +258,105 @@ void ComputeAtMap::validateAndPropagatePType() {
   }
 }
 
+void ComputeAtMap::allocateIndexVariables() {
+  // Run through all disjoint sets registered in loop map,
+  //  all lowered kir::ForLoop will correspond to one of the disjoint sets
+  //  and we only need one index variable for each set.
+  for (const auto& loop_disjoint_set : id_graph_.loopNodes().disjointSets()) {
+    ParallelType ptype;
+    // first allocate thread and grid parallel indices:
+    //  The validation pass will check that the parallel bindings within the
+    //  loop nodes are consistent so all the loops within this disjoint set
+    //  will be realized implicitly using parallel index variables.
+    if (std::any_of(
+            loop_disjoint_set->vector().begin(),
+            loop_disjoint_set->vector().end(),
+            [&ptype](IterDomain* id) {
+              if (id->isThread() &&
+                  // Halo extended parallel loops currently are handled
+                  // differently and an index variable would still
+                  // be allocated in this case.
+                  (GpuLower::current()->haloInfo().getExtent(id) == nullptr)) {
+                ptype = id->getParallelType();
+                return true;
+              }
+              return false;
+            })) {
+      loop_index_variable_map_[loop_disjoint_set.get()] =
+          NamedScalar::getParallelIndex(ptype);
+      continue;
+    }
+
+    // All loops in this set are non-parallel, non-concretized broadcast
+    //  iterdomains, their "index variable" should be zero.
+    if (std::all_of(
+            loop_disjoint_set->vector().begin(),
+            loop_disjoint_set->vector().end(),
+            [](IterDomain* id) { return id->isBroadcast(); })) {
+      loop_index_variable_map_[loop_disjoint_set.get()] = fusion_->zeroVal();
+      continue;
+    }
+
+    // Allocate variable for the iterdomains:
+    auto concrete_loop_id_it = concrete_id_cache_.find(loop_disjoint_set);
+    TORCH_INTERNAL_ASSERT(
+        concrete_loop_id_it != concrete_id_cache_.end(),
+        "Concrete id not computed");
+
+    auto concrete_loop_id = concrete_loop_id_it->second;
+
+    // Need to allocate double buffered loop differently.
+    if (GpuLower::current()->doubleBufferInfo().isDoubleBufferedIterDomain(
+            concrete_loop_id)) {
+      // Allocate index variable for each stage of the double buffered loop.
+      double_buffered_loop_index_variable_map_[loop_disjoint_set.get()] =
+          std::make_unique<DoubleBufferIndices>(DoubleBufferIndices(
+              {{DoubleBufferLoopStage::Prolog,
+                IrBuilder::create<Int>(c10::nullopt)},
+               {DoubleBufferLoopStage::Main,
+                IrBuilder::create<Int>(c10::nullopt)},
+               {DoubleBufferLoopStage::Epilog,
+                IrBuilder::create<Int>(c10::nullopt)}}));
+    } else {
+      // Everything now should be serial concrete loops,
+      //   we just allocate a loop index integer for each set of loops.
+      loop_index_variable_map_[loop_disjoint_set.get()] =
+          IrBuilder::create<Int>(c10::nullopt);
+    }
+  }
+}
+
+Val* ComputeAtMap::getIndexVariable(
+    IterDomain* id,
+    DoubleBufferLoopStage double_buffer_loop_stage) const {
+  TORCH_INTERNAL_ASSERT(
+      id_graph_.loopNodes().mappingExists(id),
+      "Index Variable: no index variable allocated as ",
+      id->toString(),
+      " is not registered in loop map");
+  const auto* loop_set = &(id_graph_.loopNodes().getDisjointSetOf(id));
+
+  // Check if this loop was modified by double buffer pass.
+  bool is_double_buffer_iterdomain =
+      GpuLower::current()->doubleBufferInfo().isDoubleBufferedIterDomain(id);
+
+  if (is_double_buffer_iterdomain) {
+    // Use dedicated double buffer index variable if the loop is double buffer
+    // loop
+    if (double_buffer_loop_stage == DoubleBufferLoopStage::NotApplicable) {
+      // The double buffered loop stages are created after the loop nest
+      //  lowering phase so this function will be querried before the double
+      //  buffer pass. At that point, no forloop has any double buffer
+      //  stage defined, and we just default to using the main stage index.
+      double_buffer_loop_stage = DoubleBufferLoopStage::Main;
+    }
+    return double_buffered_loop_index_variable_map_.at(loop_set)->at(
+        double_buffer_loop_stage);
+  } else {
+    return loop_index_variable_map_.at(loop_set);
+  }
+}
+
 bool ComputeAtMap::areMapped(
     IterDomain* id0,
     IterDomain* id1,

torch/csrc/jit/codegen/cuda/compute_at_map.h

@@ -112,6 +112,8 @@ class TORCH_CUDA_CU_API IterDomainGraph {
 
 class TrivialReductionInfo;
 
+using DoubleBufferIndices = std::unordered_map<DoubleBufferLoopStage, Int*>;
+
 class TORCH_CUDA_CU_API ComputeAtMap {
  public:
   ComputeAtMap() = delete;
@@ -122,6 +124,25 @@ class TORCH_CUDA_CU_API ComputeAtMap {
   //! all IterDomains in the disjoint set to that PType.
   void validateAndPropagatePType();
 
+  //! Run through disjoint sets in the LOOP map and allocate the index
+  //!  variable for the associated for loop that will be generated
+  //!  for each disjoint sets in the loop map. This pre-allocation makes
+  //!  2 key assumptions about computeAt map that would very likely be
+  //!  long term invariant:
+  //!    1. All kir::forloop created in the lowering pass should belong
+  //!  to one of the disjoint sets in loop map.
+  //!    2. The lowering pass will *never* create a loop nest with 2
+  //!  different nesting levels mapped together, i.e. the case below
+  //!  never occurs:
+  //!   for i in IterDomain1
+  //!    for j in IterDomain2
+  //!     ...
+  //!   With loop_map.areMapped(IterDomain1, IterDomain2) == true.
+  //! Under this condition, we can pre-allocate all required index
+  //!  variable integers before creating any kir::forloop, and this
+  //!  would help optimizing the generated integer math for indexing.
+  void allocateIndexVariables();
+
   //! Returns if id0 and id1 are mapped to eachother with provided IdMappingMode
   bool areMapped(IterDomain* id0, IterDomain* id1, IdMappingMode mode) const;
 
@@ -151,6 +172,16 @@ class TORCH_CUDA_CU_API ComputeAtMap {
   //! Get the ID sets for a provided IdMappingMode
   const DisjointSets<IterDomain*>& getIdSets(IdMappingMode mode) const;
 
+  //! Returns the pre-allocated index variable integer used in
+  //!  the kir::ForLoop corresponding to the given IterDomain.
+  //!  this interface is only valid if the ID has a loop mapping,
+  //!  ca_map will throw exceptions if given iterdomain doesn't
+  //!  have a loop map entry.
+  Val* getIndexVariable(
+      IterDomain* id,
+      DoubleBufferLoopStage double_buffer_loop_stage =
+          DoubleBufferLoopStage::NotApplicable) const;
+
  private:
   // Build id_graph_
   void build(Fusion* fusion);
@@ -178,6 +209,24 @@ class TORCH_CUDA_CU_API ComputeAtMap {
       std::shared_ptr<VectorOfUniqueEntries<IterDomain*>>,
       IterDomain*>
       concrete_id_cache_;
+
+  //! Allocated Loop index variable through the CA map.
+  //!   only valid for disjoint sets on the loop ca map.
+  std::unordered_map<const VectorOfUniqueEntries<IterDomain*>*, Val*>
+      loop_index_variable_map_;
+
+  //! Allocated loop indices for double buffer loop.
+  //!  only valid for disjoint sets on the loop ca map
+  //!  that have double buffer-ed iterdomains.
+  using DoubleBufferIndicesPtr = std::unique_ptr<DoubleBufferIndices>;
+  std::unordered_map<
+      const VectorOfUniqueEntries<IterDomain*>*,
+      DoubleBufferIndicesPtr>
+      double_buffered_loop_index_variable_map_;
+
+  // Shortcut to access the fusion this computeAt map was
+  //  built from.
+  Fusion* fusion_;
 };
 
 } // namespace cuda