perf(mge): use DepType::HOST_VALUE in trace when possible

GitOrigin-RevId: 5d47ed263fe0c5d65f86d53d61bd1c427139c06d

imperative/python/src/graph_rt.cpp

@@ -483,13 +483,13 @@ void init_graph_rt(py::module m) {
         },
         py::arg(), py::arg(), py::arg(), py::arg() = py::none(), py::arg() = py::tuple(), py::arg("graph") = py::none());
 
-    auto output_callback = [](auto callback, const std::vector<cg::VarNode*>& inputs, bool borrow = false) {
+    auto output_callback = [](auto callback, const std::vector<cg::VarNode*>& inputs, bool borrow = false, bool prefer_host_value = false) {
         SymbolVarArray sinputs;
         for (auto i : inputs) {
             sinputs.emplace_back(i);
         }
         static_assert(!std::is_reference<decltype(callback)>::value);
-        opr::OutputCallback::Param param{std::move(callback), borrow};
+        opr::OutputCallback::Param param{std::move(callback), borrow, prefer_host_value};
         auto output = opr::OutputCallback::make(std::move(param), sinputs);
         return output.node();
     };
@@ -519,7 +519,7 @@ void init_graph_rt(py::module m) {
             hv_with_event.second->record();
             p->set(std::move(hv_with_event));
         };
-        return output_callback(std::move(f), std::move(inputs), true);
+        return output_callback(std::move(f), std::move(inputs), true, true);
     });
 
     m.def("attr_output_callback", [output_callback](std::shared_ptr<Rendezvous<TensorAttr>> p, std::vector<cg::VarNode*> inputs) {

imperative/src/impl/opr_utility.cpp

@@ -144,13 +144,24 @@ cg::OperatorNodeBase::NodeProp* OutputCallback::do_make_node_prop() const {
     prop->add_flag(NodeProp::Flag::NO_AUTOMATIC_DUP);
     SmallVector<NodeProp::DepType> dep_types(input().size(),
                                              NodeProp::DepType::DEV_COMP_ORDER);
-    dep_types[0] = NodeProp::DepType::DEV_VALUE;
+    using IT = cg::static_infer::InferType;
+    auto host_value_avail = [&]() -> bool {
+        auto inp = input(0);
+        auto it = owner_graph()->static_infer_manager().get_infer_type(inp).value;
+        return it & (IT::CONST | IT::RT_STATIC | IT::MISSING_INP);
+    };
+    m_use_host_value = m_param.prefer_host_value && host_value_avail();
+    dep_types[0] = m_use_host_value ? NodeProp::DepType::HOST_VALUE : NodeProp::DepType::DEV_VALUE;
     prop->reset_dep_type(input(), dep_types);
     return prop;
 }
 
 void OutputCallback::scn_do_execute() {
-    m_param.callback(input(0)->dev_tensor());
+    if (m_use_host_value) {
+        m_param.callback(owner_graph()->static_infer_manager().infer_value(input(0)));
+    } else {
+        m_param.callback(input(0)->dev_tensor());
+    }
 }
 
 cg::OperatorNodeBase* OutputCallback::shallow_copy(

imperative/src/include/megbrain/imperative/opr_utility.h

@@ -60,7 +60,8 @@ public:
     using callback_t = thin_function<void(DeviceTensorND)>;
     struct Param {
         callback_t callback;
-        bool borrow = false;
+        bool borrow = false; // do not obtain shared ownership on DeviceTensorND
+        bool prefer_host_value = false; // use host value when possible
     };
     OutputCallback(Param param,
                    const VarNodeArray& inputs,
@@ -81,6 +82,7 @@ protected:
     NodeProp* do_make_node_prop() const override;
 private:
     Param m_param;
+    mutable bool m_use_host_value;
 };
 
 MGB_DEFINE_OPR_CLASS(NopCallback, cg::OperatorNodeBase) // {

imperative/src/test/opr_utility.cpp

@@ -13,6 +13,7 @@
 #include "megbrain/opr/io.h"
 #include "megbrain/opr/basic_arith.h"
 #include "megbrain/opr/utility.h"
+#include "megbrain/opr/tensor_manip.h"
 #include "megbrain/test/helper.h"
 
 using namespace mgb;
@@ -50,6 +51,27 @@ TEST(TestOprUtility, OutputCallback) {
     MGB_ASSERT_TENSOR_EQ(hy, *hx);
 }
 
+TEST(TestOprUtility, OutputCallbackPreferHost) {
+    HostTensorGenerator<> gen;
+    auto hx = gen({2, 3});
+    auto graph = ComputingGraph::make();
+    auto x = opr::Host2DeviceCopy::make(*graph, hx);
+    x = opr::GetVarShape::make(x);
+    HostTensorND hy;
+    auto callback = [&hy](DeviceTensorND dv) {hy.copy_from(dv);};
+    opr::OutputCallback::Param param{callback};
+    param.prefer_host_value = true;
+    auto dummy = opr::OutputCallback::make(param, x);
+    auto y = opr::VirtualDep::make({x, dummy});
+
+    ComputingGraph::OutputSpec outspec{{y, [](DeviceTensorND&){}}};
+    auto func = graph->compile(outspec);
+    func->execute();
+    ASSERT_TRUE(hy.comp_node() == CompNode::default_cpu());
+    ASSERT_EQ(hy.ptr<int>()[0], 2);
+    ASSERT_EQ(hy.ptr<int>()[1], 3);
+}
+
 TEST(TestOprUtility, NopCallback) {
     HostTensorGenerator<> gen;
     auto hx = gen({2, 3});