Move ConcretizedBroadcastDomains to shared_ptr in GpuLower. (#1988)

csarofeen · Sep 27, 2022 · 15f2f6d · 15f2f6d
1 parent 8f1c7f5
commit 15f2f6d
Show file tree

Hide file tree

Showing 8 changed files with 20 additions and 17 deletions.
diff --git a/torch/csrc/jit/codegen/cuda/lower2device.cpp b/torch/csrc/jit/codegen/cuda/lower2device.cpp
@@ -257,7 +257,8 @@ void GpuLower::lower(Fusion* fusion, DataType index_type) {
   compute_at_map_->validateAndPropagatePType();
 
   // Used in parallel dimension map
-  concretized_broadcast_domains_.build(fusion_);
+  concretized_broadcast_domains_ =
+      std::make_shared<const ConcretizedBroadcastDomains>(fusion_);
 
   parallelDimensionMap().build(fusion_);
   if (isDebugDumpEnabled(DebugDumpOption::ParallelDimensions)) {

diff --git a/torch/csrc/jit/codegen/cuda/lower2device.h b/torch/csrc/jit/codegen/cuda/lower2device.h
@@ -62,7 +62,8 @@ class TORCH_CUDA_CU_API GpuLower : public NonCopyable {
   //! Query if lowering is in progress
   static bool hasCurrent();
 
-  ConcretizedBroadcastDomains& concretizedBroadcastDomains() {
+  std::shared_ptr<const ConcretizedBroadcastDomains>
+  concretizedBroadcastDomains() {
     return concretized_broadcast_domains_;
   }
 
@@ -194,7 +195,8 @@ class TORCH_CUDA_CU_API GpuLower : public NonCopyable {
   // would be safer to wrap all of these in unique pointers and remove the build
   // interface and default constructor. That way they couldn't be accessed
   // without being initialized.
-  ConcretizedBroadcastDomains concretized_broadcast_domains_;
+  std::shared_ptr<const ConcretizedBroadcastDomains>
+      concretized_broadcast_domains_;
   ThreadPredicateMap thread_pred_map_;
   PredicateElimination pred_elimination_;
   std::shared_ptr<ComputeAtMap> compute_at_map_;

diff --git a/torch/csrc/jit/codegen/cuda/lower_sync_information.cpp b/torch/csrc/jit/codegen/cuda/lower_sync_information.cpp
@@ -26,7 +26,7 @@ void validateParallelizationOfTensor(TensorView* tv) {
     // It doesn't matter if this axis is a non-concretized broadcast
     // TODO: merging broadcast and non-broadcast
     if (axis->isBroadcast() &&
-        !GpuLower::current()->concretizedBroadcastDomains().isConcretized(
+        !GpuLower::current()->concretizedBroadcastDomains()->isConcretized(
             axis)) {
       continue;
     }
@@ -195,7 +195,7 @@ void SyncMap::build(Fusion* fusion) {
               (!parallel_bcast_doms.get(consumer_ptype) ||
                !GpuLower::current()
                     ->concretizedBroadcastDomains()
-                    .isConcretized(consumer_axis))) {
+                    ->isConcretized(consumer_axis))) {
             continue;
           }
 
@@ -421,7 +421,7 @@ void SyncMap::build(Fusion* fusion) {
                                      .redundant_types;
 
           if (p_id->isBroadcast() &&
-              GpuLower::current()->concretizedBroadcastDomains().isConcretized(
+              GpuLower::current()->concretizedBroadcastDomains()->isConcretized(
                   p_id) &&
               producer->getMemoryType() == MemoryType::Shared &&
               redundant_preds.hasTID()) {
@@ -436,7 +436,7 @@ void SyncMap::build(Fusion* fusion) {
               (!parallel_bcast_doms.get(producer_ptype) ||
                !GpuLower::current()
                     ->concretizedBroadcastDomains()
-                    .isConcretized(p_id))) {
+                    ->isConcretized(p_id))) {
             continue;
           }
 

diff --git a/torch/csrc/jit/codegen/cuda/lower_thread_predicate.cpp b/torch/csrc/jit/codegen/cuda/lower_thread_predicate.cpp
@@ -237,7 +237,7 @@ void ThreadPredicateMap::updateBitSet(const Expr* expr) {
           id_reductions.set(id->getParallelType());
         }
         if (id->isBroadcast() &&
-            GpuLower::current()->concretizedBroadcastDomains().isConcretized(
+            GpuLower::current()->concretizedBroadcastDomains()->isConcretized(
                 id)) {
           id_bcasts.set(id->getParallelType());
         }
@@ -575,7 +575,8 @@ ParallelTypeBitmap ThreadPredicateMap::getParallelBroadcastDomains(
 
   for (auto id : iter_domains) {
     if (!id->isBroadcast() ||
-        !GpuLower::current()->concretizedBroadcastDomains().isConcretized(id)) {
+        !GpuLower::current()->concretizedBroadcastDomains()->isConcretized(
+            id)) {
       continue;
     }
     if (id->isBlockDim() || (!output_smem && id->isThreadDim())) {

diff --git a/torch/csrc/jit/codegen/cuda/lower_trivial_broadcast.cpp b/torch/csrc/jit/codegen/cuda/lower_trivial_broadcast.cpp
@@ -10,7 +10,7 @@ namespace jit {
 namespace fuser {
 namespace cuda {
 
-void ConcretizedBroadcastDomains::build(Fusion* fusion) {
+ConcretizedBroadcastDomains::ConcretizedBroadcastDomains(Fusion* fusion) {
   exact_map_ = std::make_unique<ExactRootDomainMap>(fusion);
 
   // Initialize the origin map with input broadcast domains

diff --git a/torch/csrc/jit/codegen/cuda/lower_trivial_broadcast.h b/torch/csrc/jit/codegen/cuda/lower_trivial_broadcast.h
@@ -23,7 +23,8 @@ namespace cuda {
 //! domains are marked as concretized.
 class TORCH_CUDA_CU_API ConcretizedBroadcastDomains : private IterVisitor {
  public:
-  void build(Fusion* fusion);
+  ConcretizedBroadcastDomains() = delete;
+  ConcretizedBroadcastDomains(Fusion* fusion);
 
   //! Is a domain concretized?
   bool isConcretized(IterDomain* id) const;

diff --git a/torch/csrc/jit/codegen/cuda/scheduler/registry.cpp b/torch/csrc/jit/codegen/cuda/scheduler/registry.cpp
@@ -789,8 +789,7 @@ static bool checkPatternEquivalence(
 // being broadcasted to one size multiple times or different sizes. This is a
 // hard to optimize problem and likely indicates we shouldn't be fusing.
 bool hasNonUniqueBcast(Fusion* fusion) {
-  ConcretizedBroadcastDomains concretize_info;
-  concretize_info.build(fusion);
+  ConcretizedBroadcastDomains concretize_info(fusion);
 
   for (auto tv : ir_utils::allTvs(fusion)) {
     for (auto id : tv->getRootDomain()) {

diff --git a/torch/csrc/jit/codegen/cuda/test/test_gpu.cpp b/torch/csrc/jit/codegen/cuda/test/test_gpu.cpp
@@ -20889,9 +20889,9 @@ TEST_F(NVFuserTest, FusionBroadcastConcretization1_CUDA) {
   }
 
   GpuLower gpulw(&fusion);
-  TORCH_CHECK(!gpulw.concretizedBroadcastDomains().isConcretized(
+  TORCH_CHECK(!gpulw.concretizedBroadcastDomains()->isConcretized(
       loweredTv(tv4, gpulw)->axis(1)));
-  TORCH_CHECK(gpulw.concretizedBroadcastDomains().isConcretized(
+  TORCH_CHECK(gpulw.concretizedBroadcastDomains()->isConcretized(
       loweredTv(tv7, gpulw)->axis(1)));
 
   auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);
@@ -21079,8 +21079,7 @@ TEST_F(NVFuserTest, FusionBroadcastConcretization5_CUDA) {
   auto tvs3 = Welford(tv17, {1});
   fusion.addOutput(tvs3.avg);
 
-  ConcretizedBroadcastDomains bcast_concretization_info;
-  bcast_concretization_info.build(&fusion);
+  ConcretizedBroadcastDomains bcast_concretization_info(&fusion);
 
   TORCH_CHECK(
       bcast_concretization_info.maybeNonUniquelyConcretized(tv5->axis(1)),