apache · vinx13 · Jul 13, 2021 · Jun 30, 2021 · Jun 30, 2021 · Jun 30, 2021
diff --git a/docs/dev/inferbound.rst b/docs/dev/inferbound.rst
@@ -447,13 +447,11 @@ Here is the IR after ScheduleOps (note that loops with extent 1 have been preser
 
 ::
 
-   // attr [compute(D, 0x2c070b0)] realize_scope = ""
    realize D([0, 4], [0, 5], [0, 16]) {
      produce D {
        for (di, 0, 4) {
          for (dj, 0, 5) {
            for (dk, 0, 16) {
-             // attr [compute(C, 0x2c29990)] realize_scope = ""
              realize C([dj, 1], [dk, 1]) {
                produce C {
                  for (i, 0, 1) {

diff --git a/include/tvm/te/operation.h b/include/tvm/te/operation.h
@@ -125,11 +125,12 @@ class TVM_DLL OperationNode : public Object {
    * \param stage the op's stage.
    * \param realize_map The realization domain map of the operators.
    * \param body The body that is going to get
+   * \param storage_scope The storage scope associated with this realization
    * \return A realization statement that wraps body.
    */
   virtual Stmt BuildRealize(const Stage& stage,
-                            const std::unordered_map<IterVar, Range>& realize_map,
-                            const Stmt& body) const = 0;
+                            const std::unordered_map<IterVar, Range>& realize_map, const Stmt& body,
+                            String storage_scope = "") const = 0;
   /*!
    * \brief Build the statement that provide the output tensors.
    * \param stage The schedule stage of the op.
@@ -168,7 +169,7 @@ class PlaceholderOpNode : public OperationNode {
   void GatherBound(const Operation& self, const std::unordered_map<Tensor, TensorDom>& tensor_dom,
                    std::unordered_map<IterVar, Range>* out_dom_map) const final;
   Stmt BuildRealize(const Stage& stage, const std::unordered_map<IterVar, Range>& realize_map,
-                    const Stmt& body) const final;
+                    const Stmt& body, String storage_scope = "") const final;
   Stmt BuildProvide(const Stage& stage, const std::unordered_map<IterVar, Range>& dom_map,
                     bool debug_keep_trivial_loop) const final;
 
@@ -212,7 +213,7 @@ class TVM_DLL BaseComputeOpNode : public OperationNode {
   void GatherBound(const Operation& self, const std::unordered_map<Tensor, TensorDom>& tensor_dom,
                    std::unordered_map<IterVar, Range>* out_dom_map) const final;
   Stmt BuildRealize(const Stage& stage, const std::unordered_map<IterVar, Range>& realize_map,
-                    const Stmt& body) const final;
+                    const Stmt& body, String storage_scope = "") const final;
   virtual size_t num_schedulable_dims() const = 0;
 
   static constexpr const char* _type_key = "BaseComputeOp";
@@ -370,7 +371,7 @@ class ScanOpNode : public OperationNode {
   void GatherBound(const Operation& self, const std::unordered_map<Tensor, TensorDom>& tensor_dom,
                    std::unordered_map<IterVar, Range>* out_dom_map) const final;
   Stmt BuildRealize(const Stage& stage, const std::unordered_map<IterVar, Range>& realize_map,
-                    const Stmt& body) const final;
+                    const Stmt& body, String storage_scope = "") const final;
   Stmt BuildProvide(const Stage& stage, const std::unordered_map<IterVar, Range>& dom_map,
                     bool debug_keep_trivial_loop) const final;
 
@@ -433,7 +434,7 @@ class ExternOpNode : public OperationNode {
   void GatherBound(const Operation& self, const std::unordered_map<Tensor, TensorDom>& tensor_dom,
                    std::unordered_map<IterVar, Range>* out_dom_map) const final;
   Stmt BuildRealize(const Stage& stage, const std::unordered_map<IterVar, Range>& realize_map,
-                    const Stmt& body) const final;
+                    const Stmt& body, String storage_scope = "") const final;
   Stmt BuildProvide(const Stage& stage, const std::unordered_map<IterVar, Range>& dom_map,
                     bool debug_keep_trivial_loop) const final;
 
@@ -498,7 +499,7 @@ class HybridOpNode : public OperationNode {
   void GatherBound(const Operation& self, const std::unordered_map<Tensor, TensorDom>& tensor_dom,
                    std::unordered_map<IterVar, Range>* out_dom_map) const final;
   Stmt BuildRealize(const Stage& stage, const std::unordered_map<IterVar, Range>& realize_map,
-                    const Stmt& body) const final;
+                    const Stmt& body, String storage_scope = "") const final;
   Stmt BuildProvide(const Stage& stage, const std::unordered_map<IterVar, Range>& dom_map,
                     bool debug_keep_trivial_loop) const final;
 

diff --git a/include/tvm/tir/buffer.h b/include/tvm/tir/buffer.h
@@ -191,12 +191,13 @@ class Buffer : public ObjectRef {
  * \param shape The shape of the buffer,
  * \param dtype The content data type.
  * \param name The name of the buffer
+ * \param storage_scope The storage scope associated with this buffer
  * \param span The location of this object in the source code.
  * \return The created buffer.
  * \sa Buffer for complete constructor.
  */
 TVM_DLL Buffer decl_buffer(Array<PrimExpr> shape, DataType dtype = DataType::Float(32),
-                           String name = "buffer", Span span = Span());
+                           String name = "buffer", String storage_scope = "", Span span = Span());
 
 /*!
  * \brief Base node for data producers.

diff --git a/include/tvm/tir/stmt.h b/include/tvm/tir/stmt.h
@@ -464,25 +464,30 @@ class ProducerRealizeNode : public StmtNode {
   PrimExpr condition;
   /*! \brief The body of realization. */
   Stmt body;
+  /*! \brief The storage scope associated with this realization. */
+  String storage_scope;
 
   void VisitAttrs(AttrVisitor* v) {
     v->Visit("producer", &producer);
     v->Visit("bounds", &bounds);
     v->Visit("condition", &condition);
     v->Visit("body", &body);
+    v->Visit("storage_scope", &storage_scope);
     v->Visit("span", &span);
   }
 
   bool SEqualReduce(const ProducerRealizeNode* other, SEqualReducer equal) const {
     return equal(producer, other->producer) && equal(bounds, other->bounds) &&
-           equal(condition, other->condition) && equal(body, other->body);
+           equal(condition, other->condition) && equal(body, other->body) &&
+           equal(storage_scope, other->storage_scope);
   }
 
   void SHashReduce(SHashReducer hash_reduce) const {
     hash_reduce(producer);
     hash_reduce(bounds);
     hash_reduce(condition);
     hash_reduce(body);
+    hash_reduce(storage_scope);
   }
 
   static constexpr const char* _type_key = "tir.ProducerRealize";
@@ -496,7 +501,7 @@ class ProducerRealizeNode : public StmtNode {
 class ProducerRealize : public Stmt {
  public:
   TVM_DLL ProducerRealize(DataProducer producer, Region bounds, PrimExpr condition, Stmt body,
-                          Span span = Span());
+                          String storage_scope = "", Span span = Span());
 
   TVM_DEFINE_OBJECT_REF_METHODS(ProducerRealize, Stmt, ProducerRealizeNode);
 };

diff --git a/python/tvm/script/scope_handler.py b/python/tvm/script/scope_handler.py
@@ -140,7 +140,7 @@ def enter_scope(
 
         def setup_buffer_var(extents, dtype, scope, condition=True, span: Span = None):
             """Setup buffer var for a given type."""
-            buffer_ptr_type = tvm.ir.PointerType(tvm.ir.PrimType(dtype))
+            buffer_ptr_type = tvm.ir.PointerType(tvm.ir.PrimType(dtype), scope)
             self.buffer_var = tvm.tir.Var(name, buffer_ptr_type, span)
 
         setup_buffer_var(*arg_list, span=tvm_span_from_synr(var_span))

diff --git a/python/tvm/script/special_stmt.py b/python/tvm/script/special_stmt.py
@@ -491,6 +491,22 @@ def var(dtype, span):
         super().__init__(var, def_symbol=True)
 
 
+@register
+class BufferVarDef(SpecialStmt):
+    """Special function for defining a variable of pointer type"""
+
+    def __init__(self):
+        def buffer_var(dtype, storage_scope, span):
+            assert isinstance(
+                self.node, ast.Assign
+            ), f"BufferVarDef expected ast.Assign but got {type(self.node)}"
+            ptr_type = tvm.ir.PointerType(tvm.ir.PrimType(dtype), storage_scope)
+            v = te.var(self.node.lhs.id.name, ptr_type, span=span)
+            self.context.update_symbol(v.name, v, self.node)
+
+        super().__init__(buffer_var, def_symbol=True)
+
+
 @register
 class EnvThread(SpecialStmt):
     """Bind a var to thread env"""

diff --git a/python/tvm/te/hybrid/parser.py b/python/tvm/te/hybrid/parser.py
@@ -207,8 +207,7 @@ def wrap_up_realize(self, node, body):
             _domain = [Range.from_min_extent(0, i) for i in _buf.shape]
             _dtype = _buf.dtype
             _true = tvm.runtime.convert(True)
-            body = tvm.tir.ProducerRealize(_buf, _domain, _true, body)
-            body = tvm.tir.AttrStmt(_buf.op, "realize_scope", tvm.runtime.convert(_scope), body)
+            body = tvm.tir.ProducerRealize(_buf, _domain, _true, body, tvm.runtime.convert(_scope))
 
         for elem in to_pop:
             self.symbols.pop(elem)

diff --git a/python/tvm/tir/buffer.py b/python/tvm/tir/buffer.py
@@ -250,7 +250,7 @@ def decl_buffer(
         # Bool is represented as uint1 in the IR, but stored as int8
         storage_type = PrimType(dtype)
         storage_type = PrimType("int8") if storage_type.dtype == "bool" else storage_type
-        data = Var(name, PointerType(storage_type), span)
+        data = Var(name, PointerType(storage_type, scope), span)
     return _ffi_api.Buffer(
         data,
         dtype,

diff --git a/python/tvm/tir/ir_builder.py b/python/tvm/tir/ir_builder.py
@@ -394,7 +394,7 @@ def let(self, var_name, value):
         self.emit(lambda x: _stmt.LetStmt(var, value, x))
         return var
 
-    def allocate(self, dtype, shape, name="buf", scope=None):
+    def allocate(self, dtype, shape, name="buf", scope=""):
         """Create a allocate statement.
 
         Parameters
@@ -416,15 +416,15 @@ def allocate(self, dtype, shape, name="buf", scope=None):
         buffer : BufferVar
             The buffer var representing the buffer.
         """
-        buffer_var = _expr.Var(name, PointerType(PrimType(dtype)))
+        buffer_var = _expr.Var(name, PointerType(PrimType(dtype), scope))
         if not isinstance(shape, (list, tuple, _container.Array)):
             shape = [shape]
         if scope:
             self.scope_attr(buffer_var, "storage_scope", scope)
         self.emit(lambda x: _stmt.Allocate(buffer_var, dtype, shape, const(1, dtype="uint1"), x))
         return BufferVar(self, buffer_var, shape, dtype)
 
-    def pointer(self, content_type, name="ptr"):
+    def pointer(self, content_type, name="ptr", scope=""):
         """Create pointer variable with content type.
 
         Parameters
@@ -435,12 +435,15 @@ def pointer(self, content_type, name="ptr"):
         name : str, optional
             The name of the pointer.
 
+        scope : str, optional
+            The scope of the pointer.
+
         Returns
         -------
         ptr : BufferVar
             The buffer var representing the buffer.
         """
-        buffer_var = _expr.Var(name, dtype="handle")
+        buffer_var = _expr.Var(name, PointerType(PrimType(content_type), scope))
         return BufferVar(self, buffer_var, None, content_type)
 
     def buffer_ptr(self, buf, shape=None):

diff --git a/python/tvm/tir/stmt.py b/python/tvm/tir/stmt.py
@@ -364,13 +364,16 @@ class ProducerRealize(Stmt):
     body : Stmt
         The realize body
 
+    storage_scope : str
+        The storage scope associated with this realization
+
     span : Optional[Span]
         The location of this itervar in the source code.
     """
 
-    def __init__(self, producer, bounds, condition, body, span=None):
+    def __init__(self, producer, bounds, condition, body, storage_scope="", span=None):
         self.__init_handle_by_constructor__(
-            _ffi_api.ProducerRealize, producer, bounds, condition, body, span
+            _ffi_api.ProducerRealize, producer, bounds, condition, body, storage_scope, span
         )
 
 

diff --git a/src/contrib/hybrid/codegen_hybrid.cc b/src/contrib/hybrid/codegen_hybrid.cc
@@ -315,10 +315,6 @@ void CodeGenHybrid::VisitStmt_(const AttrStmtNode* op) {
     indent_ += tab_;
     PrintStmt(op->body);
     indent_ -= tab_;
-  } else if (op->attr_key == tir::attr::realize_scope) {
-    auto v = Downcast<Operation>(op->node);
-    alloc_storage_scope_[v] = op->value.as<StringImmNode>()->value;
-    PrintStmt(op->body);
   } else {
     // For now we ignore the unsupported AttrStmt
     PrintStmt(op->body);
@@ -327,8 +323,7 @@ void CodeGenHybrid::VisitStmt_(const AttrStmtNode* op) {
 
 void CodeGenHybrid::VisitStmt_(const ProducerRealizeNode* op) {
   auto tensor = Downcast<Tensor>(op->producer);
-  ICHECK(alloc_storage_scope_.count(tensor->op));
-  if (!alloc_storage_scope_[tensor->op].empty()) {
+  if (!op->storage_scope.empty()) {
     PrintIndent();
     stream << GetTensorID(tensor) << " = allocate((";
     for (size_t i = 0; i < op->bounds.size(); ++i) {
@@ -339,7 +334,7 @@ void CodeGenHybrid::VisitStmt_(const ProducerRealizeNode* op) {
     stream << "), '";
     PrintType(tensor->dtype, stream);
     stream << "', '";
-    stream << alloc_storage_scope_[tensor->op] << "')\n";
+    stream << op->storage_scope << "')\n";
   }
   PrintStmt(op->body);
 }

diff --git a/src/contrib/hybrid/codegen_hybrid.h b/src/contrib/hybrid/codegen_hybrid.h
@@ -168,8 +168,6 @@ class CodeGenHybrid : public ExprFunctor<void(const PrimExpr&, std::ostream&)>,
    * \param tensor The tensor to allocate a name.
    */
   std::string GetTensorID(const Tensor& tensor);
-  /*! \brief the storage scope of allocation */
-  std::map<Operation, std::string> alloc_storage_scope_;
 };
 
 }  // namespace contrib

diff --git a/src/printer/tvmscript_printer.cc b/src/printer/tvmscript_printer.cc
@@ -26,6 +26,7 @@
 #include <tvm/ir/module.h>
 #include <tvm/node/serialization.h>
 #include <tvm/runtime/registry.h>
+#include <tvm/tir/buffer.h>
 #include <tvm/tir/expr.h>
 #include <tvm/tir/expr_functor.h>
 #include <tvm/tir/function.h>
@@ -1013,8 +1014,17 @@ Doc TVMScriptPrinter::PrintPrimFunc(const PrimFunc& primFunc) {
       return memo_var_[GetRef<Var>(a)].str() < memo_var_[GetRef<Var>(b)].str();
     });
     for (const auto& var : vars) {
-      header_var << Doc::NewLine() << Print(GetRef<Var>(var)) << " = tir.var(";
-      header_var << PrintDType(var->dtype) << ")";
+      auto type = GetRef<Var>(var)->type_annotation;
+      if (auto* ptr_type = type.as<PointerTypeNode>()) {
+        auto* prim_type = ptr_type->element_type.as<PrimTypeNode>();
+        ICHECK(prim_type);
+        header_var << Doc::NewLine() << Print(GetRef<Var>(var)) << " = tir.buffer_var(";
+        header_var << PrintDType(prim_type->dtype) << ", "
+                   << Doc::StrLiteral(ptr_type->storage_scope) << ")";
+      } else {
+        header_var << Doc::NewLine() << Print(GetRef<Var>(var)) << " = tir.var(";
+        header_var << PrintDType(var->dtype) << ")";
+      }
     }
   }
   doc << Doc::Indent(4, header_attr << header_var << header_buf << body);

diff --git a/src/relay/backend/aot_executor_codegen.cc b/src/relay/backend/aot_executor_codegen.cc
@@ -722,7 +722,8 @@ class AOTExecutorCodegen : public ExprVisitor {
     // Define the storage allocator ids
     for (auto kv : storage_device_map_) {
       for (auto sid : kv.second->storage_ids) {
-        te::Var buffer_var(MakeString("sid_", sid), PointerType(PrimType(DataType::Int(8))));
+        te::Var buffer_var(MakeString("sid_", sid),
+                           PointerType(PrimType(DataType::Int(8)), "global"));
         sids_table_[sid] = buffer_var;
       }
     }

diff --git a/src/runtime/thread_storage_scope.h b/src/runtime/thread_storage_scope.h
@@ -118,7 +118,9 @@ struct StorageScope {
    */
   static StorageScope Create(const std::string& s) {
     StorageScope r;
-    if (s.compare(0, 6, "global") == 0) {
+    if (s.empty()) {
+      r.rank = StorageRank::kGlobal;
+    } else if (s.compare(0, 6, "global") == 0) {
       r.rank = StorageRank::kGlobal;
       r.tag = s.substr(6, std::string::npos);
     } else if (s.compare(0, 6, "shared") == 0) {

diff --git a/src/target/llvm/codegen_amdgpu.cc b/src/target/llvm/codegen_amdgpu.cc
@@ -84,7 +84,8 @@ class CodeGenAMDGPU : public CodeGenLLVM {
     if (info.alignment > 16) {
       info.alignment = 16;
     }
-    if (info.scope.rank == runtime::StorageRank::kLocal) {
+    auto storage_scope = runtime::StorageScope::Create(GetPtrStorageScope(op->buffer_var));
+    if (storage_scope.rank == runtime::StorageRank::kLocal) {
       // const int local_address_space = 5;
       // TODO(tqchen): for higher version of LLVM, local address space can be set.
       llvm::AllocaInst* alloca = WithFunctionEntry([&]() {
@@ -99,7 +100,7 @@ class CodeGenAMDGPU : public CodeGenLLVM {
       }
       buf = alloca;
     } else {
-      ICHECK(info.scope.rank == runtime::StorageRank::kShared)
+      ICHECK(storage_scope.rank == runtime::StorageRank::kShared)
           << "Can only allocate shared or local memory inside kernel";
       // Shared memory: address space  == 3
       const unsigned shared_address_space = 3;

diff --git a/src/target/llvm/codegen_llvm.cc b/src/target/llvm/codegen_llvm.cc
@@ -501,7 +501,8 @@ void CodeGenLLVM::GetAlignment(DataType t, const VarNode* buf_var, const PrimExp
   auto it = alloc_storage_info_.find(buf_var);
   if (it != alloc_storage_info_.end()) {
     const StorageInfo& info = it->second;
-    *p_native_bits = NativeVectorBits(info.scope);
+    *p_native_bits =
+        NativeVectorBits(runtime::StorageScope::Create(GetPtrStorageScope(GetRef<Var>(buf_var))));
     max_align_bits = info.alignment * 8;
   } else {
     *p_native_bits = native_vector_bits_;
@@ -1390,11 +1391,6 @@ void CodeGenLLVM::VisitStmt_(const AttrStmtNode* op) {
         analyzer_->Bind(iv->var, Range::FromMinExtent(0, op->value));
       }
     }
-  } else if (op->attr_key == tir::attr::storage_scope) {
-    const VarNode* v = op->node.as<VarNode>();
-    ICHECK(v);
-    alloc_storage_info_[v].scope =
-        runtime::StorageScope::Create(op->value.as<StringImmNode>()->value);
   } else if (op->attr_key == tir::attr::storage_alignment) {
     const VarNode* v = op->node.as<VarNode>();
     ICHECK(v);

diff --git a/src/target/llvm/codegen_llvm.h b/src/target/llvm/codegen_llvm.h
@@ -163,8 +163,6 @@ class CodeGenLLVM : public ExprFunctor<llvm::Value*(const PrimExpr&)>,
  protected:
   /*! \brief The storage information */
   struct StorageInfo {
-    /*! \brief The storage scope */
-    runtime::StorageScope scope;
     /*! \brief The alignment of allocation */
     int alignment{0};
   };