refactor(mgb): replace static_cast<ComputingGraphImpl*> with a checked version

GitOrigin-RevId: d05b114668f84f9b4045abe55af11bce0ff3bd4a

CMakeLists.txt

@@ -224,6 +224,7 @@ endif()
 option(MGE_WITH_DISTRIBUTED "Build with distributed support" ON)
 option(MGE_BUILD_XXX "Build _xxx.so instead of mgb.so " OFF)
 if(MGE_BUILD_XXX)
+    add_compile_definitions(MGB_ENABLE_IMPERATIVE_RUNTIME)
     set(CMAKE_CXX_STANDARD 17)
 endif()
 

src/core/impl/graph/cg_impl.cpp

@@ -674,7 +674,7 @@ void ComputingGraphImpl::share_device_memory_with(ComputingGraph& other) {
     mgb_assert(
             !m_current_comp_seq,
             "share_device_memory_with must be called before compiling graph");
-    auto&& oimpl = static_cast<ComputingGraphImpl&>(other);
+    auto&& oimpl = *ComputingGraphImpl::downcast(&other);
     var_node_mem_manager().static_device_memory_manager(
             oimpl.var_node_mem_manager().static_device_memory_manager());
 }
@@ -707,7 +707,7 @@ size_t ComputingGraphImpl::clear_device_memory() {
 }
 
 void ComputingGraphImpl::set_as_subgraph(ComputingGraph& par_graph) {
-    m_parent_graph = static_cast<ComputingGraphImpl*>(&par_graph);
+    m_parent_graph = ComputingGraphImpl::downcast(&par_graph);
     m_parent_graph->m_subgraphs.emplace_back(this);
     m_node_id_counter = m_parent_graph->m_node_id_counter;
     options().var_sanity_check_first_run =

src/core/impl/graph/cg_impl.h

@@ -122,6 +122,15 @@ public:
     ComputingGraphImpl();
     ~ComputingGraphImpl();
 
+    template<typename T> static ComputingGraphImpl* downcast(T* ptr) = delete;
+
+    inline static ComputingGraphImpl* downcast(ComputingGraph* graph) {
+        #ifdef MGB_ENABLE_IMPERATIVE_RUNTIME
+        mgb_assert(!graph->options().imperative_proxy_graph);
+        #endif
+        return static_cast<ComputingGraphImpl*>(graph);
+    }
+
     friend struct ComputingGraph::Options;
 
     std::unique_ptr<AsyncExecutable> compile(

src/core/impl/graph/cg_impl_seq.h

@@ -100,7 +100,7 @@ class ComputingGraphImpl::ComputingSequence final : public AsyncExecutable {
 public:
     ComputingSequence(const std::shared_ptr<ComputingGraph>& graph)
             : m_owner_graph_refkeep{graph},
-              m_owner_graph{static_cast<ComputingGraphImpl*>(graph.get())},
+              m_owner_graph{ComputingGraphImpl::downcast(graph.get())},
               m_have_parent_graph{m_owner_graph->m_parent_graph} {}
 
     GraphExecutable::ExecEnv& exec_env() { return m_exec_env; }

src/core/impl/graph/eager_eval.cpp

@@ -76,7 +76,7 @@ EagerEvalManager::~EagerEvalManager() noexcept {
     if (m_first_opr_enable_status == 1) {
         m_var_sync_mgr_pool.disable_freelist();
         for (auto&& i :
-             static_cast<ComputingGraphImpl*>(m_owner_graph)->all_oprs()) {
+             ComputingGraphImpl::downcast(m_owner_graph)->all_oprs()) {
             for (auto var : i->output()) {
                 auto mgr = VarNodeMemManager::var_node_cn_sync_manager(var);
                 if (mgr) {
@@ -223,7 +223,7 @@ void EagerEvalManager::prepare_for_exec(OperatorNodeBase* opr) {
 }
 
 void EagerEvalManager::update_static_infer_result(OperatorNodeBase* opr) {
-    auto&& mgr = static_cast<ComputingGraphImpl*>(m_owner_graph)
+    auto&& mgr = ComputingGraphImpl::downcast(m_owner_graph)
                          ->static_infer_manager_impl();
     auto sync_missing_trait =
             [&](static_infer::StaticInferManagerImpl::TagHandler* handler) {
@@ -260,7 +260,7 @@ void EagerEvalManager::update_static_infer_result(OperatorNodeBase* opr) {
 }
 
 void EagerEvalManager::ensure_input_layout(VarNode* var) {
-    auto&& mem_mgr = static_cast<ComputingGraphImpl*>(var->owner_graph())
+    auto&& mem_mgr = ComputingGraphImpl::downcast(var->owner_graph())
                              ->var_node_mem_manager();
 
     auto trait = mem_mgr.get_var_node_mem_trait_nullable(var);
@@ -287,7 +287,7 @@ void EagerEvalManager::alloc_output_mem(OperatorNodeBase* opr) {
         }
     }
 
-    auto&& mgr = static_cast<ComputingGraphImpl*>(m_owner_graph)
+    auto&& mgr = ComputingGraphImpl::downcast(m_owner_graph)
                          ->var_node_mem_manager();
     OprNodeArray opr_seq{opr};
 
@@ -348,7 +348,7 @@ void EagerEvalManager::do_on_opr_insert(OperatorNodeBase* opr) {
         if (status) {
             update_static_infer_result(opr);
             alloc_output_mem(opr);
-            auto&& mgr = static_cast<ComputingGraphImpl*>(m_owner_graph)
+            auto&& mgr = ComputingGraphImpl::downcast(m_owner_graph)
                              ->var_node_mem_manager();
             mgr.on_graph_compile_finished();
             opr->execute(*m_exec_env);

src/core/impl/graph/grad_manager.cpp

@@ -40,7 +40,7 @@ class GradShapeChecker {
 
     void do_on_var_shape(VarNode *var) {
         MGB_MARK_USED_VAR(m_opr);
-        auto graph = static_cast<ComputingGraphImpl*>(var->owner_graph());
+        auto graph = ComputingGraphImpl::downcast(var->owner_graph());
 
         auto seq = graph->current_comp_seq();
         if (seq) {
@@ -90,7 +90,7 @@ class GradShapeChecker {
         }
 
         static void make(OperatorNodeBase *opr, VarNode *wrt, VarNode *grad) {
-            if (static_cast<ComputingGraphImpl*>(wrt->owner_graph())
+            if (ComputingGraphImpl::downcast(wrt->owner_graph())
                     ->eager_eval_manager().enabled())
                 return;
             using namespace std::placeholders;
@@ -650,13 +650,13 @@ void GradManager::add_var_virtual_receiver(
 }
 
 void cg::add_grad_transformer(VarNode *var, const GradTransformer &cb) {
-    static_cast<ComputingGraphImpl*>(var->owner_graph())->
+    ComputingGraphImpl::downcast(var->owner_graph())->
         grad_manager().
         add_grad_transformer(var, cb);
 }
 
 void cg::add_extra_dep_for_grad(VarNode *inp, VarNode *out) {
-    static_cast<ComputingGraphImpl*>(inp->owner_graph())->grad_manager().
+    ComputingGraphImpl::downcast(inp->owner_graph())->grad_manager().
         add_extra_dep_for_grad(inp, out);
 }
 
@@ -667,7 +667,7 @@ void cg::add_var_virtual_receiver(
     desc->inputs = inputs;
     desc->outputs = outputs;
     desc->grad = grad;
-    static_cast<ComputingGraphImpl*>(inputs.at(0)->owner_graph())->
+    ComputingGraphImpl::downcast(inputs.at(0)->owner_graph())->
         grad_manager().
         add_var_virtual_receiver(desc);
 }

src/core/impl/graph/helper.cpp

@@ -99,8 +99,8 @@ SymbolVarArray cg::grad(SymbolVar target_, SymbolVarArray wrts_, bool warn_mid_w
     grads.reserve(wrts_.size());
     VarNodeArray dest_vars;
     auto&& graph = target->owner_graph();
-    auto&& eager_mgr = static_cast<ComputingGraphImpl*>(graph)->eager_eval_manager();
-    auto&& grad_mgr = static_cast<ComputingGraphImpl*>(graph)->grad_manager();
+    auto&& eager_mgr = ComputingGraphImpl::downcast(graph)->eager_eval_manager();
+    auto&& grad_mgr = ComputingGraphImpl::downcast(graph)->grad_manager();
     bool already_recorded = eager_mgr.enter_record_mode();
     for (auto&& wrt_ : wrts_) {
         auto wrt = wrt_.node();
@@ -139,7 +139,7 @@ SymbolVarArray cg::grad(SymbolVar target_, SymbolVarArray wrts_, bool warn_mid_w
 
 SymbolVar cg::current_grad_target(ComputingGraph &graph) {
 #if MGB_ENABLE_GRAD
-    auto var = static_cast<ComputingGraphImpl&>(graph).grad_manager(
+    auto var = ComputingGraphImpl::downcast(&graph)->grad_manager(
             ).current_grad_target();
     mgb_throw_if(!var, GraphError, "current_grad_target() called outside "
             "grad computing environment");

src/core/impl/graph/operator_node.cpp

@@ -93,7 +93,7 @@ OperatorNodeBase::OperatorNodeBase(ComputingGraph *owner,
 }
 
 OperatorNodeBase::~OperatorNodeBase() noexcept {
-    auto &&pool = static_cast<ComputingGraphImpl*>(
+    auto &&pool = ComputingGraphImpl::cast(
             owner_graph())->var_node_pool();
     for (auto i: m_output) {
         pool.free(i);
@@ -124,7 +124,7 @@ void OperatorNodeBase::execute(ExecEnv &env) {
     }
 
     // allocate output with dynamic storage
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
             ->var_node_mem_manager()
             .alloc_var_node_mem_dynamic(env, this);
 
@@ -135,11 +135,11 @@ void OperatorNodeBase::execute(ExecEnv &env) {
     // static_infer_manager so the value would be up-to-date; however for shape
     // deps, oprs would access the shape directly, so we need to insert some
     // code here to ensure it is up-to-date.
-    if (!static_cast<ComputingGraphImpl*>(owner_graph())
+    if (!ComputingGraphImpl::downcast(owner_graph())
                  ->eager_eval_manager()
                  .enabled()) {
         VarNodeArray vars_to_set;
-        auto cg = static_cast<ComputingGraphImpl*>(owner_graph());
+        auto cg = ComputingGraphImpl::downcast(owner_graph());
         auto step_cur = cg->opr_step_num_in_cur_comp_seq(this).val();
         mgb_assert(step_cur < std::numeric_limits<size_t>::max());
         using DT = NodeProp::DepType;
@@ -264,7 +264,7 @@ VarNode* OperatorNodeBase::add_output(const Maybe<std::string> &name) {
     mgb_assert(!m_inserted_in_graph && !m_node_prop.valid(),
             "add output on opr after it has been inserted into graph");
 
-    auto ptr = static_cast<ComputingGraphImpl*>(
+    auto ptr = ComputingGraphImpl::cast(
                 owner_graph())->var_node_pool().alloc(
                 name.valid() ? this->name() + ":" + name.val() : name, this);
     m_output.push_back(ptr);
@@ -676,7 +676,7 @@ void mixin::IOSameShapeOperatorNode::get_output_var_shape(
 
 void PostExecActions::add(VarNode* var) {
     mgb_assert(m_comp_node == var->comp_node());
-    auto graph = static_cast<ComputingGraphImpl*>(var->owner_graph());
+    auto graph = ComputingGraphImpl::downcast(var->owner_graph());
 
     auto&& infer_mgr = graph->static_infer_manager_impl();
     auto&& extra_info = graph->current_comp_seq_extra_info();

src/core/impl/graph/static_infer_impl.cpp

@@ -813,7 +813,7 @@ StaticInferManagerImpl::~StaticInferManagerImpl() noexcept {
     m_mem_pool_value_trait.disable_freelist();
     for (auto &&i: m_dtor_callbacks)
         i.second();
-    for (auto &&i: static_cast<ComputingGraphImpl*>(
+    for (auto &&i: ComputingGraphImpl::downcast(
                 m_owner_graph)->all_oprs()) {
         for (auto j: i->output()) {
             clear_tag_handler(j);
@@ -1212,7 +1212,7 @@ class StaticInferManagerImpl::SubgraphStaticInferHelperImpl final:
 
     void check_graph_par(VarNode *var) {
         if (mgb_unlikely(!m_par_graph)) {
-            m_par_graph = static_cast<ComputingGraphImpl*>(var->owner_graph());
+            m_par_graph = ComputingGraphImpl::downcast(var->owner_graph());
             mgb_assert(m_par_graph != m_sub_graph);
 
             auto cb = [this]() {
@@ -1230,7 +1230,7 @@ class StaticInferManagerImpl::SubgraphStaticInferHelperImpl final:
 
     void check_graph_sub(VarNode *var) {
         if (mgb_unlikely(!m_sub_graph)) {
-            m_sub_graph = static_cast<ComputingGraphImpl*>(var->owner_graph());
+            m_sub_graph = ComputingGraphImpl::downcast(var->owner_graph());
             mgb_assert(m_sub_graph != m_par_graph);
         } else {
             mgb_assert(m_sub_graph == var->owner_graph());

src/core/impl/graph/symbol_var.cpp

@@ -132,7 +132,7 @@ const DeviceTensorND& SymbolVar::eager_eval_get_value() const {
 #if MGB_BUILD_SLIM_SERVING
     mgb_throw(MegBrainError, "eager eval disabled at compile time");
 #else
-    auto og = static_cast<ComputingGraphImpl*>(node()->owner_graph());
+    auto og = ComputingGraphImpl::downcast(node()->owner_graph());
     mgb_assert(og->options().eager_evaluation);
     return node()->dev_tensor();
 #endif

src/core/impl/graph/topo_sort.cpp

@@ -260,7 +260,7 @@ void TopoSorter::DFSDepDiscover::proc_add_dep_comp_order1() {
 void TopoSorter::DFSDepDiscover::proc_find_missing_inp() {
     auto frame = m_cur_frame;
     auto opr = frame->opr;
-    auto&& mgr = static_cast<ComputingGraphImpl*>(opr->owner_graph())
+    auto&& mgr = ComputingGraphImpl::downcast(opr->owner_graph())
                          ->static_infer_manager_impl();
     auto&& missing_inp = frame->missing_inputs;
 

src/core/impl/graph/var_node.cpp

@@ -233,12 +233,12 @@ bool VarNode::set_fwd_in2out_readonly(
     if (owner_graph()->options().imperative_proxy_graph) {
         return false;
     }
-    return static_cast<ComputingGraphImpl*>(owner_graph())
+    return ComputingGraphImpl::downcast(owner_graph())
         ->var_node_mem_manager().fwd_in2out_readonly(input, sub, this);
 }
 
 VarNode& VarNode::set_fwd_in2out_writable(VarNode *input) {
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
         ->var_node_mem_manager().fwd_in2out_writable(input, this);
     return *this;
 }
@@ -246,20 +246,20 @@ VarNode& VarNode::set_fwd_in2out_writable(VarNode *input) {
 
 VarNode& VarNode::set_fwd_in2out_writable_force(VarNode *input) {
     mgb_assert(!owner_graph()->options().imperative_proxy_graph);
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
         ->var_node_mem_manager().fwd_in2out_writable_force(input, this);
     return *this;
 }
 
 VarNode& VarNode::add_layout_constraint(LayoutConstraintCallback callback) {
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
         ->var_node_mem_manager().add_layout_constraint(
                 this, std::move(callback));
     return *this;
 }
 
 VarNode& VarNode::add_layout_constraint_contiguous() {
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
             ->var_node_mem_manager()
             .add_layout_constraint_level(
                     this, VarNodeMemManager::LayoutConstraintLevel::CONTIG);
@@ -267,7 +267,7 @@ VarNode& VarNode::add_layout_constraint_contiguous() {
 }
 
 VarNode& VarNode::add_layout_constraint_monotone() {
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
             ->var_node_mem_manager()
             .add_layout_constraint_level(
                     this, VarNodeMemManager::LayoutConstraintLevel::MONOTONE);
@@ -315,7 +315,7 @@ VarNode& VarNode::shape_alloc(const TensorShape &shape) {
                 "shape_alloc() could only be used for vars with"
                 " NO_SYS_MEM_ALLOC flag; actual var: %s",
                 cg::dump_var_info({this}).c_str());
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
         ->var_node_mem_manager().var_alloc_with_shape(this, shape);
     return *this;
 }
@@ -330,7 +330,7 @@ bool VarNode::reset_dev_tensor_from_other_var(VarNode* src_var) {
                 "dynamic storage on src is required for dynamic readonly "
                 "forwarding: vars=%s",
                 dump_var_info({src_var, this}).c_str());
-        auto&& trait = static_cast<ComputingGraphImpl*>(owner_graph())
+        auto&& trait = ComputingGraphImpl::downcast(owner_graph())
                                ->var_node_mem_manager()
                                .get_var_node_mem_trait_at(src_var);
         if (trait.seq_force_update_dest ||
@@ -403,7 +403,7 @@ std::shared_ptr<json::Value> VarNode::to_json() const {
         return json::Null::make();
     };
 
-    auto &&trait = static_cast<ComputingGraphImpl*>(owner_graph()
+    auto &&trait = ComputingGraphImpl::downcast(owner_graph()
             )->var_node_mem_manager().get_var_node_mem_trait(this);
     auto flag = json::Array::make();
     {
@@ -459,7 +459,7 @@ std::shared_ptr<json::Value> VarNode::to_json() const {
 #endif
 
 MemAllocPlan& VarNode::init_mem_plan(const DeviceTensorND* fixed_alloc) {
-    static_cast<ComputingGraphImpl*>(owner_graph())
+    ComputingGraphImpl::downcast(owner_graph())
             ->var_node_mem_manager()
             .init_single_var_mem_plan(this, fixed_alloc);
     return m_mem_plan;
@@ -477,7 +477,7 @@ void VarNode::modify_flag(Flag delta, Flag new_flag) {
                     Flag::NO_SYS_STATIC_MEM_ALLOC |
                     Flag::RT_FORCE_DYNAMIC_MEM_ALLOC)) == delta);
 
-        mgb_assert(!static_cast<ComputingGraphImpl*>(owner_graph())->
+        mgb_assert(!ComputingGraphImpl::downcast(owner_graph())->
                 var_node_mem_manager().optimize_started(),
                 "could not modify var flags after optimization started");
     }

src/core/impl/graph/var_node_mem_mgr.cpp

@@ -340,7 +340,7 @@ VarNodeMemManager::DynamicAllocOprInfo::DynamicAllocOprInfo(
     prev_dev_val_input.clear();
     static_infer_inp.clear();
     dev_val_input.clear();
-    auto &&mgr = static_cast<ComputingGraphImpl*>(opr->owner_graph())->
+    auto &&mgr = ComputingGraphImpl::downcast(opr->owner_graph())->
         static_infer_manager_impl();
 
     CompNode single_cn;

src/core/impl/graph/var_node_mem_mgr/defrag.cpp

@@ -73,7 +73,7 @@ void VarDevMemDefragmenter::defrag(VarNode* req_var,
                                    const CompNodeInfo& cn_info,
                                    size_t extra_size) {
     // pause all other comp nodes before calling defrag_impl()
-    auto exec_env = static_cast<ComputingGraphImpl*>(req_var->owner_graph())
+    auto exec_env = ComputingGraphImpl::downcast(req_var->owner_graph())
                             ->current_exec_env();
     mgb_assert(exec_env);
     exec_env->pause_exec();