shared_ptr<IrContainer> transition and Fusion tracking infrastructure

Change Fusion::ir_container_ from unique_ptr to shared_ptr to enable
future container sharing between Fusions. Add Fusion tracking API to
IrContainer (addFusion/removeFusion/transferFusion/sharingCount).
Remove IrContainer::parent_ since the 1:1 relationship no longer holds.
Disable parallel compilation during the shared_ptr transition.

Author: mdavis36

csrc/fusion.cpp

@@ -117,8 +117,6 @@ void Fusion::swap(Fusion& a, Fusion& b) noexcept {
   // update the parent backpointers in those containers to point to their new
   // owners
   if (a.ir_container_) {
-    // Also update all Statement ir_container_ pointers to point to new owner
-    a.ir_container()->parent_ = &a;
     for (auto val : a.vals()) {
       val->ir_container_ = &a;
     }
@@ -127,8 +125,6 @@ void Fusion::swap(Fusion& a, Fusion& b) noexcept {
     }
   }
   if (b.ir_container_) {
-    // Also update all Statement ir_container_ pointers to point to new owner
-    b.ir_container()->parent_ = &b;
     for (auto val : b.vals()) {
       val->ir_container_ = &b;
     }
@@ -162,7 +158,8 @@ std::unique_ptr<SegmentedFusion> Fusion::segment(
 IrCloner Fusion::copy(const Fusion* from, Fusion* to) {
   to->clear();
 
-  auto ir_cloner = IrContainer::copy(from->ir_container(), to->ir_container());
+  auto ir_cloner =
+      IrContainer::copy(from->ir_container(), to->ir_container(), to);
 
   // Remap cached special val pointers through the cloner
   if (from->zero_val_) {
@@ -255,8 +252,8 @@ IrCloner Fusion::copy(const Fusion* from, Fusion* to) {
 }
 
 // Default constructor
-Fusion::Fusion() : ir_container_(std::make_unique<IrContainer>()) {
-  ir_container_->parent_ = this;
+Fusion::Fusion() : ir_container_(std::make_shared<IrContainer>()) {
+  ir_container_->addFusion(this);
 }
 
 // Copy constructor
@@ -288,6 +285,9 @@ Fusion& Fusion::operator=(Fusion&& other) noexcept {
 
 Fusion::~Fusion() {
   clear();
+  if (ir_container_) {
+    ir_container_->removeFusion(this);
+  }
 }
 
 void Fusion::clear() noexcept {
@@ -351,9 +351,7 @@ void Fusion::removeExpr(Expr* expr) {
   auto expr_in_deque = std::find_if(
       c->exprs_up_.begin(),
       c->exprs_up_.end(),
-      [expr](std::unique_ptr<Expr>& expr_up) {
-        return expr_up.get() == expr;
-      });
+      [expr](std::unique_ptr<Expr>& expr_up) { return expr_up.get() == expr; });
   NVF_ERROR(
       expr_in_deque != c->exprs_up_.end(),
       "Wanted to remove an expression but its unique ptr is missing.");

csrc/fusion.h

@@ -149,7 +149,6 @@ class NVF_API Fusion : public PolymorphicBase {
   typedef std::unordered_map<int, std::vector<int64_t>> PermutationMap;
 
  protected:
-  // Direct access to underlying container
   IrContainer* ir_container() {
     NVF_ERROR(
         ir_container_.get() != nullptr,
@@ -164,6 +163,10 @@ class NVF_API Fusion : public PolymorphicBase {
     return ir_container_.get();
   }
 
+  std::shared_ptr<IrContainer> ir_container_ptr() const {
+    return ir_container_;
+  }
+
  public:
   // Registration (public API with passkey)
   virtual void registerStmt(IrBuilderPasskey, Statement* stmt) {
@@ -636,7 +639,7 @@ class NVF_API Fusion : public PolymorphicBase {
   std::unique_ptr<std::vector<TensorView*>> all_tvs_ptr_ = nullptr;
 
   inline static const std::string exact_mappings_key = "exact_mappings";
-  std::unique_ptr<IrContainer> ir_container_;
+  std::shared_ptr<IrContainer> ir_container_;
 
   Val* zero_val_ = nullptr;
   Val* one_val_ = nullptr;

csrc/ir/container.cpp

@@ -80,14 +80,15 @@ void IrContainer::swap(IrContainer& a, IrContainer& b) noexcept {
 
   std::swap(a.val_type_name_map_, b.val_type_name_map_);
   std::swap(a.expr_name_counter_, b.expr_name_counter_);
-
-  std::swap(a.parent_, b.parent_);
 }
 
-IrCloner IrContainer::copy(const IrContainer* from, IrContainer* to) {
+IrCloner IrContainer::copy(
+    const IrContainer* from,
+    IrContainer* to,
+    Fusion* dest_fusion) {
   to->clear();
 
-  IrCloner ir_cloner(to->parent());
+  IrCloner ir_cloner(dest_fusion);
 
   // Copy values in deterministic order
   for (auto val : from->deterministic_vals()) {
@@ -138,7 +139,7 @@ bool IrContainer::inContainer(const Statement* const_stmt) const {
   }
 
   NVF_ERROR(
-      const_stmt->container() == this->parent(),
+      sharing_fusions_.count(const_stmt->container()) > 0,
       "Container claims to own stmt, but stmt disagrees.");
 
   // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
@@ -157,4 +158,29 @@ bool IrContainer::inContainer(const Statement* const_stmt) const {
   return true;
 }
 
+void IrContainer::addFusion(Fusion* fusion) {
+  sharing_fusions_.insert(fusion);
+}
+
+void IrContainer::removeFusion(Fusion* fusion) {
+  sharing_fusions_.erase(fusion);
+}
+
+void IrContainer::transferFusion(Fusion* from, Fusion* to) {
+  sharing_fusions_.erase(from);
+  sharing_fusions_.insert(to);
+}
+
+size_t IrContainer::sharingCount() const {
+  return sharing_fusions_.size();
+}
+
+bool IrContainer::hasMultipleFusions() const {
+  return sharing_fusions_.size() > 1;
+}
+
+const std::unordered_set<Fusion*>& IrContainer::sharingFusions() const {
+  return sharing_fusions_;
+}
+
 } // namespace nvfuser

csrc/ir/container.h

@@ -86,7 +86,10 @@ class IrContainer {
   }
 
  protected:
-  static IrCloner copy(const IrContainer* from, IrContainer* to);
+  static IrCloner copy(
+      const IrContainer* from,
+      IrContainer* to,
+      Fusion* dest_fusion);
 
   static void swap(IrContainer& a, IrContainer& b) noexcept;
 
@@ -127,16 +130,15 @@ class IrContainer {
   StmtNameType expr_name_counter_ = 0;
 
  public:
-  Fusion* parent() const {
-    NVF_ERROR(
-        parent_ != nullptr, "Call to IrContainer::parent() holds nullptr.")
-    return parent_;
-  }
+  void addFusion(Fusion* fusion);
+  void removeFusion(Fusion* fusion);
+  void transferFusion(Fusion* from, Fusion* to);
+  size_t sharingCount() const;
+  bool hasMultipleFusions() const;
+  const std::unordered_set<Fusion*>& sharingFusions() const;
 
  private:
-  // Parent Fusion that owns this container (for pure composition pattern)
-  // Used by Statement::fusion() to navigate back to owning Fusion
-  Fusion* parent_ = nullptr;
+  std::unordered_set<Fusion*> sharing_fusions_;
 };
 
 } // namespace nvfuser

csrc/runtime/fusion_kernel_runtime.cpp

@@ -26,6 +26,9 @@
 
 namespace nvfuser {
 
+// TODO: Remove when std::shared_mutex is added to IrContainer.
+constexpr bool kPhase2DisableParallelCompile = true;
+
 namespace {
 // Replace CUDA tensor with Meta tensor because storing tensors can cause
 // out-of-memory issues. Other arguments are returned as-is.
@@ -436,7 +439,8 @@ void FusionKernelRuntime::compileFusionParallel(KernelArgumentHolder args) {
   try {
     for (const auto& [group_to_run, group_runtime_inputs] :
          zip(runtime_workspace_.group_run_order, all_runtime_inputs)) {
-      if (num_groups == 1 || isOptionDisabled(DisableOption::ParallelCompile)) {
+      if (num_groups == 1 || kPhase2DisableParallelCompile ||
+          isOptionDisabled(DisableOption::ParallelCompile)) {
         compileKernel(group_runtime_inputs, group_to_run);
       } else {
         // launch compileKernel thread here
@@ -470,7 +474,8 @@ void FusionKernelRuntime::compileFusionParallel(KernelArgumentHolder args) {
     throw;
   }
 
-  if (num_groups != 1 && !isOptionDisabled(DisableOption::ParallelCompile)) {
+  if (num_groups != 1 && !kPhase2DisableParallelCompile &&
+      !isOptionDisabled(DisableOption::ParallelCompile)) {
     // Wait until all segments finish compiling
     getThreadPool()->waitWorkComplete();
     NVF_ERROR(