feat(mge/profiler): add compatible and graphviz mode

GitOrigin-RevId: 40ca9ea80e

imperative/python/megengine/utils/profiler.py

@@ -9,13 +9,155 @@
 import base64
 import json
 import os
-from typing import List, Optional
+import re
+from typing import Iterable, List, Optional
 
 from ..core._imperative_rt import OperatorNodeConfig, ProfileEntry
 from ..core._imperative_rt import ProfilerImpl as _Profiler
 from ..core._imperative_rt.imperative import sync
 from ..core._imperative_rt.ops import CollectiveCommMode
-from ..core.ops.builtin import GetVarShape
+
+
+def _make_dict(**kwargs):
+    unused_keys = []
+    for k, v in kwargs.items():
+        if v is None:
+            unused_keys.append(k)
+    for k in unused_keys:
+        del kwargs[k]
+    return kwargs
+
+
+def _print_opnode_config(config):
+    return _make_dict(
+        name=config.name, dtype=config.dtype, comp_node_arr=config.comp_node_arr,
+    )
+
+
+def _dump_chrome_timeline(entries: List[ProfileEntry], path: str):
+    pid = os.getpid()
+    trace_events = []
+
+    def append_event(**kwargs):
+        trace_events.append(_make_dict(**kwargs))
+
+    for id, entry in enumerate(entries):
+        op = entry.op
+        name = type(op).__name__
+        host_begin, host_end = entry.host
+        device_list = entry.device_list
+        args = Profiler.fetch_attrs(op)
+        args["__id__"] = "[{}]".format(id)
+        cat = name
+        for ts, ph in [(host_begin, "B"), (host_end, "E")]:
+            append_event(
+                name=name, ph=ph, ts=ts * 1000, pid=pid, tid="host", args=args, cat=cat,
+            )
+        for device, device_begin, device_end in device_list:
+            for ts, ph in [(device_begin(), "B"), (device_end(), "E")]:
+                append_event(
+                    name=name, ph=ph, ts=ts * 1000, pid=pid, tid=str(device), args=args,
+                )
+    with open("{}.chrome_timeline.json".format(path), "w") as f:
+        json.dump(trace_events, f, indent=2)
+
+
+def _dump_compatible(entries: List[ProfileEntry], path: str):
+    obj = {
+        "graph_exec": {"var": [], "operator": {}},
+        "profiler": {"device": {}, "host": {}, "opr_footprint": {}},
+    }
+    var_list = obj["graph_exec"]["var"]
+    operator_dict = obj["graph_exec"]["operator"]
+    device_dict = obj["profiler"]["device"]
+    host_dict = obj["profiler"]["host"]
+    opr_foot_print_dict = obj["profiler"]["opr_footprint"]
+
+    def add_var(var) -> int:
+        var_id = len(var_list)
+        var_list.append(
+            {"comp_node": str(var[2]),}
+        )
+        return var_id
+
+    for op_id, entry in enumerate(entries):
+        operator_dict[op_id] = {
+            "input": [add_var(var) for var in entry.inputs],
+            "output": [add_var(var) for var in entry.outputs],
+            "name": str(entry.op.ctype()),
+            "type": "imperative",
+            "id": entry.id,
+        }
+        op_device_dict = {}
+        for device, device_begin, device_end in entry.device_list:
+            op_device_dict[str(device)] = {
+                "start": device_begin(),
+                "kern": device_begin(),
+                "end": device_end(),
+            }
+        device_dict[op_id] = op_device_dict
+        host_begin, host_end = entry.host
+        host_dict[op_id] = {
+            "host": {"start": host_begin, "kern": host_begin, "end": host_end}
+        }
+        opr_footprint = {
+            "out_shapes": [oup[1] for oup in entry.outputs],
+            "in_shapes": [inp[1] for inp in entry.inputs],
+            "params": {},
+        }
+        if entry.memory > 0:
+            opr_footprint["memory"] = entry.memory
+        if entry.computation > 0:
+            opr_footprint["computation"] = entry.computation
+        opr_foot_print_dict[op_id] = opr_footprint
+    with open("{}.compatible.json".format(path), "w") as f:
+        json.dump(obj, f, indent=2)
+
+
+def _dump_graphviz(entries: List[ProfileEntry], path: str):
+    import graphviz
+    import json
+
+    graph = graphviz.Digraph()
+    graph.graph_attr["ordering"] = "out"
+    var_cache = {}
+
+    def cache_var(var_id, var_shape):
+        if var_id not in var_cache:
+            var_name = "var({})".format(var_id)
+            var_label = "{}\nshape:{}\n".format(var_name, shape)
+            graph.node(var_name, var_label)
+            var_cache[var_id] = var_name
+        return var_cache[var_id]
+
+    for op_id, entry in enumerate(entries):
+        op = entry.op
+        op_name = "op({})".format(op_id)
+        op_type = type(op).__name__
+        op_attrs = Profiler.fetch_attrs(op)
+        label_lines = []
+        if "param" in op_attrs:
+            del op_attrs["param"]
+        label_lines.append("{}:{}".format(op_name, op_type))
+        for k, v in op_attrs.items():
+            label_lines.append("attr[{}]: {}".format(k, v))
+        op_param_str = entry.param
+        if len(op_param_str) > 0:
+            op_param = json.loads(op_param_str)
+            for k, v in op_param.items():
+                label_lines.append("param[{}]:{}".format(k, v))
+        host_begin, host_end = entry.host
+        label_lines.append("time[host]: {:f}ms".format(host_end - host_begin))
+        for device, device_begin, device_end in entry.device_list:
+            device_time = device_end() - device_begin()
+            label_lines.append("time[{}]: {:f}ms".format(device, device_time))
+        op_label = "\n".join(label_lines)
+        graph.node(op_name, op_label, shape="rectangle")
+        for var_id, shape, device in entry.inputs:
+            graph.edge(cache_var(var_id, shape), op_name)
+        for var_id, shape, device in entry.outputs:
+            graph.edge(op_name, cache_var(var_id, shape))
+    graph.save("{}.graphviz.dot".format(path))
 
 
 class Profiler:
@@ -23,7 +165,7 @@ class Profiler:
     Profile graph execution in imperative mode.
 
     :type path: Optional[str]
-    :param path: default path for profiler to dump.
+    :param path: default path prefix for profiler to dump.
 
     Examples:
 
@@ -31,59 +173,67 @@ class Profiler:
 
         import megengine as mge
         import megengine.module as M
-        import megengine.utils.profiler.Profiler
+        from megengine.utils.profiler import Profiler
 
         # With Learnable Parameters
         for iter in range(0, 10):
             # Only profile record of last iter would be saved
-            with Profiler("profile.json"):
+            with Profiler("profile"):
                 # your code here
         
         # Then open the profile file in chrome timeline window
     """
 
-    # see https://github.com/catapult-project/catapult/blob/master/tracing/tracing/base/color_scheme.html
-    GOOD = "good"
-    BAD = "bad"
-    TERRIBLE = "terrible"
+    CHROME_TIMELINE = "chrome_timeline"
+    COMPATIBLE = "compatible"
+    GRAPHVIZ = "graphviz"
+
+    WITH_FOOTPRINT = 1
 
-    BLACK = "black"
-    GREY = "grey"
-    WHITE = "white"
-    YELLOW = "yellow"
-    OLIVE = "olive"
+    _type_map = {
+        OperatorNodeConfig: lambda x: _print_opnode_config(x),
+        bytes: lambda x: base64.encodebytes(x).decode("ascii"),
+        CollectiveCommMode: lambda x: str(x),
+    }
 
-    def __init__(self, path: str = "profile.json"):
+    _dumper_map = {
+        CHROME_TIMELINE: _dump_chrome_timeline,
+        COMPATIBLE: _dump_compatible,
+        GRAPHVIZ: _dump_graphviz,
+    }
+
+    def __init__(
+        self,
+        path: str = "profile",
+        *,
+        formats: Iterable[str] = (CHROME_TIMELINE,),
+        type_filter: str = ".*",
+        exit_dump: bool = True
+    ) -> None:
         self._impl = _Profiler()
         self._path = path
-        self._color_map = {}
-        self._type_map = {
-            OperatorNodeConfig: lambda x: self.print_opnode_config(x),
-            bytes: lambda x: base64.encodebytes(x).decode("ascii"),
-            CollectiveCommMode: lambda x: str(x),
-        }
+
+        if isinstance(formats, str):
+            formats = (formats,)
+
+        self._filter = type_filter
+        self._dumpers = [Profiler._dumper_map[fmt] for fmt in formats]
+        self._exit_dump = exit_dump
 
     def __enter__(self):
         sync()
-        self._impl.start()
+        self._impl.start(Profiler.WITH_FOOTPRINT)
         return self
 
-    def __exit__(self, val, type, trace):
+    def __exit__(self, val, tp, trace):
+        if self._exit_dump:
+            self.dump()
         sync()
         self._impl.stop()
-        if self._path is not None:
-            self.dump()
-
-    def recolor(self, target: str, color: str):
-        self._color_map[target] = color
-        return self
+        self._impl.clear()
 
-    def print_opnode_config(self, config):
-        return self.make_dict(
-            name=config.name, dtype=config.dtype, comp_node_arr=config.comp_node_arr,
-        )
-
-    def fetch_attrs(self, op):
+    @classmethod
+    def fetch_attrs(cls, op):
         attrs = dir(op)
         results = {}
         for attr in attrs:
@@ -93,61 +243,29 @@ class Profiler:
             if callable(value):
                 continue
             value_type = type(value)
-            if value_type in self._type_map:
-                value = self._type_map[value_type](value)
+            if value_type in cls._type_map:
+                value = cls._type_map[value_type](value)
             results[attr] = value
         return results
 
-    def make_dict(self, **kwargs):
-        unused_keys = []
-        for k, v in kwargs.items():
-            if v is None:
-                unused_keys.append(k)
-        for k in unused_keys:
-            del kwargs[k]
-        return kwargs
-
     def dump(self, path: Optional[str] = None):
-        pid = os.getpid()
+        sync()
+        raw = [
+            entry
+            for entry in self._impl.dump()
+            if re.match(self._filter, type(entry.op).__name__)
+        ]
         if path is None:
             path = self._path
-        trace_events = []
-
-        def append_event(**kwargs):
-            trace_events.append(self.make_dict(**kwargs))
-
-        entries: List[ProfileEntry] = self._impl.dump()
-
-        for id, entry in enumerate(entries):
-            op = entry.op
-            name = type(op).__name__
-            host_begin, host_end = entry.host
-            device_list = entry.device_list
-            args = self.fetch_attrs(op)
-            args["__id__"] = "[{}]".format(id)
-            cname = self._color_map[name] if name in self._color_map else None
-            cat = name
-            for ts, ph in [(host_begin, "B"), (host_end, "E")]:
-                append_event(
-                    name=name,
-                    ph=ph,
-                    ts=ts * 1000,
-                    pid=pid,
-                    tid="host",
-                    args=args,
-                    cname=cname,
-                    cat=cat,
-                )
-            for device, device_begin, device_end in device_list:
-                for ts, ph in [(device_begin(), "B"), (device_end(), "E")]:
-                    append_event(
-                        name=name,
-                        ph=ph,
-                        ts=ts * 1000,
-                        pid=pid,
-                        tid=str(device),
-                        args=args,
-                        cname=cname,
-                    )
-        with open(path, "w") as f:
-            json.dump(trace_events, f, indent=2)
+        for dumper in self._dumpers:
+            dumper(raw, path)
+
+    def __call__(self, func):
+        def wrapper(*args, **kwargs):
+            with self:
+                return func(*args, **kwargs)
+
+        return wrapper
+
+
+profile = Profiler

imperative/python/src/utils.cpp

@@ -204,17 +204,27 @@ void init_utils(py::module m) {
     py::class_<ProfileEntry>(m, "ProfileEntry")
             .def_readwrite("op", &ProfileEntry::op)
             .def_readwrite("host", &ProfileEntry::host)
-            .def_readwrite("device_list", &ProfileEntry::device_list);
+            .def_readwrite("device_list", &ProfileEntry::device_list)
+            .def_readwrite("inputs", &ProfileEntry::inputs)
+            .def_readwrite("outputs", &ProfileEntry::outputs)
+            .def_readwrite("id", &ProfileEntry::id)
+            .def_readwrite("parent", &ProfileEntry::parent)
+            .def_readwrite("memory", &ProfileEntry::memory)
+            .def_readwrite("computation", &ProfileEntry::computation)
+            .def_property_readonly("param", [](ProfileEntry& self)->std::string{
+                if(self.param){
+                    return self.param->to_string();
+                } else {
+                    return {};
+                }
+            });
 
     py::class_<mgb::imperative::Profiler>(m, "ProfilerImpl")
             .def(py::init<>())
-            .def("start",
-                 [](mgb::imperative::Profiler& profiler) { profiler.start(); })
-            .def("stop",
-                 [](mgb::imperative::Profiler& profiler) { profiler.stop(); })
-            .def("dump", [](mgb::imperative::Profiler& profiler) {
-                return profiler.get_profile();
-            });
+            .def("start", &mgb::imperative::Profiler::start)
+            .def("stop", &mgb::imperative::Profiler::stop)
+            .def("clear", &mgb::imperative::Profiler::clear)
+            .def("dump", &mgb::imperative::Profiler::get_profile);
 
     using mgb::imperative::TensorSanityCheck;
     py::class_<TensorSanityCheck>(m, "TensorSanityCheckImpl")

imperative/src/include/megbrain/imperative/function_hook.h → imperative/src/impl/function_hook.h

@@ -15,6 +15,7 @@
 
 namespace mgb {
 namespace imperative {
+
 template <typename TFunction>
 class FunctionHooker;
 
@@ -22,13 +23,18 @@ template <typename TRet, typename... TArgs>
 class FunctionHooker<TRet(TArgs...)> {
 public:
     using FunctionType = thin_function<TRet(TArgs&&...)>;
+    //Type of hooks. Hook should accept a real function as argument
+    //and invoke it on an appropriate time
     using HookType = thin_function<TRet(FunctionType, TArgs&&...)>;
-    explicit FunctionHooker(FunctionType* fptr) : m_fptr{fptr} {}
+    explicit FunctionHooker(FunctionType* fptr) : m_fptr{fptr} {
+        m_backup = {nullptr, [](FunctionType*){}};
+    }
 
 public:
     FunctionHooker& apply_hook(HookType&& hook) {
         if (!m_backup) {
             FunctionType* backup = new FunctionType(*m_fptr);
+            //Restore hooked function, would be invoked when destructed
             std::function<void(FunctionType*)> restorer =
                     [fptr = m_fptr](FunctionType* bkp) -> void {
                 *fptr = *bkp;
@@ -36,9 +42,11 @@ public:
             };
             m_backup = decltype(m_backup)(backup, restorer);
         }
+        //Replace with hooked version
         *m_fptr = [func = *m_fptr, hook](TArgs&&... args) -> TRet {
             return hook(func, std::forward<TArgs>(args)...);
         };
+        //Convinent for chain call
         return *this;
     }
 
@@ -47,9 +55,15 @@ private:
     std::unique_ptr<FunctionType, std::function<void(FunctionType*)>> m_backup;
 };
 
+//Helps to deduce template args
 template <typename TRet, typename... TArgs>
 FunctionHooker(thin_function<TRet(TArgs...)>* f)
         ->FunctionHooker<TRet(TArgs...)>;
-}  // namespace imperative
 
+template<typename TSignature>
+auto make_shared_hook(thin_function<TSignature>* fptr){
+    return std::make_shared<FunctionHooker<TSignature>>(fptr);
+}
+
+}  // namespace imperative
 }  // namespace mgb

imperative/src/impl/profiler.cpp

@@ -11,19 +11,20 @@
 
 #include "megbrain/imperative/profiler.h"
 
-#include <variant>
-
+#include "./function_hook.h"
 #include "megbrain/imperative/ops/opr_attr.h"
 #include "megbrain/imperative/physical_tensor.h"
 
+#include "megbrain/plugin/opr_footprint.h"
+
 #include "./event_pool.h"
 #include "./op_trait.h"
 
 namespace mgb {
-
 namespace imperative {
 
 namespace {
+
 CompNode::UnorderedSet collect_comp_nodes(
         const OpDef& def, const SmallVector<TensorPtr>& inputs) {
     CompNode::UnorderedSet comp_nodes;
@@ -36,37 +37,101 @@ CompNode::UnorderedSet collect_comp_nodes(
     return comp_nodes;
 }
 
+DeviceTimer::SharedEvent alloc_recorded_event(CompNode device) {
+    auto event = EventPool::with_timer().alloc_shared(device);
+    event->record();
+    return event;
+}
+
+OprFootprint footprint{};
+
 }  // namespace
 
 void DeviceTimer::reset(thin_function<double()> host_timer) {
     CompNode::foreach ([this, host_timer](CompNode device) {
-        auto base_event = EventPool::with_timer().alloc_shared(device);
-        base_event->record();
-        m_base_event_table[device] = {std::move(base_event), host_timer()};
+        m_base_event_table[device] = {alloc_recorded_event(device), host_timer()};
     });
+    m_host_timer = host_timer;
 }
 
 thin_function<double()> DeviceTimer::get_device_time(CompNode device) {
     auto event = EventPool::with_timer().alloc_shared(device);
     event->record();
+    if(m_base_event_table.count(device) == 0) {
+        m_base_event_table[device] = {alloc_recorded_event(device), m_host_timer()};
+    }
     auto base = m_base_event_table[device];
     return [base, event] {
         auto [base_event, host_time] = base;
-        //TODO: sync once for each compnode
+        // TODO: sync once for each compnode
         event->host_wait();
         return base_event->elapsed_time_until(*event) * 1000 + host_time;
     };
 }
 
-void Profiler::start() {
+void DeviceTimer::clear() {
+    m_base_event_table.clear();
+}
+
+size_t TensorRecorder::record_tensor(const TensorPtr& tensor) {
+    if (m_tensor_map.count(tensor.get()) > 0) {
+        auto& [prev, id] = m_tensor_map[tensor.get()];
+        if (prev.lock() != tensor) {
+            prev = tensor;
+            id = m_next_id++;
+        }
+        return id;
+    } else {
+        auto id = m_next_id++;
+        m_tensor_map.insert({tensor.get(), {std::weak_ptr{tensor}, id}});
+        return id;
+    }
+}
+
+void TensorRecorder::clear() {
+    m_next_id = 0;
+    m_tensor_map.clear();
+}
+
+Profile& Profiler::get_profile() {
+    for (auto& entry : m_profile) {
+        for (auto& [device, device_begin, device_end] : entry.device_list) {
+            MGB_MARK_USED_VAR(device);
+            device_begin = [value = device_begin()] { return value; };
+            device_end = [value = device_end()] { return value; };
+        }
+    }
+    return m_profile;
+}
+
+void Profiler::start(uint32_t flags) {
     m_host_timer.reset();
-    m_device_timer.reset([&]{ return m_host_timer.get_msecs();} );
-    OpTrait::for_each_trait([this](OpTrait& trait) {
-        FunctionHooker hooker{&trait.apply_on_physical_tensor};
-        hooker.apply_hook([this](auto&& apply, const OpDef& def,
-                                 const SmallVector<TensorPtr>& inputs) {
+    m_device_timer.reset([&] { return m_host_timer.get_msecs(); });
+    OpTrait::for_each_trait([this, flags](OpTrait& trait) {
+        auto hook_apply_on_physical_tensor =
+                make_shared_hook(&trait.apply_on_physical_tensor);
+        auto hook_apply_on_var_node =
+                make_shared_hook(&trait.apply_on_var_node);
+        hook_apply_on_physical_tensor->apply_hook([this, flags]
+                (auto&& apply, const OpDef& def, const SmallVector<TensorPtr>& inputs) {
+            auto shape2vector = [](const TensorShape& shape) {
+                std::vector<size_t> vector_shape;
+                for (size_t i = 0; i < shape.ndim; i++) {
+                    vector_shape.push_back(shape[i]);
+                }
+                return vector_shape;
+            };
             ProfileEntry entry;
+            entry.id = m_entry_count++;
+            // TODO: assign parent
+            entry.parent = 0;
+            // Record apply context and save to m_profile
             entry.op = def.copy();
+            for (auto&& input : inputs) {
+                entry.inputs.push_back({m_tensor_recorder.record_tensor(input),
+                                        shape2vector(input->layout()),
+                                        input->comp_node()});
+            }
             double host_begin = m_host_timer.get_msecs();
             auto&& comp_nodes = collect_comp_nodes(def, inputs);
             for (auto&& comp_node : comp_nodes) {
@@ -75,6 +140,11 @@ void Profiler::start() {
                          m_device_timer.get_device_time(comp_node),
                          {}});
             }
+            if (flags & PROFILE_FOOTPRINT) {
+                MGB_LOCK_GUARD(m_lock);
+                m_entry_stack.push({&def, &entry, std::this_thread::get_id()});
+            }
+            // Do real apply
             auto outputs = apply(def, inputs);
             for (auto& [cn, dev_begin, dev_end] : entry.device_list) {
                 MGB_MARK_USED_VAR(cn);
@@ -82,20 +152,71 @@ void Profiler::start() {
                 dev_end = m_device_timer.get_device_time(cn);
             }
             entry.host = {host_begin, m_host_timer.get_msecs()};
-            m_profile->push_back(std::move(entry));
+            for (auto&& output : outputs) {
+                entry.outputs.push_back(
+                        {m_tensor_recorder.record_tensor(output),
+                         shape2vector(output->layout()), output->comp_node()});
+            }
+            if (flags & PROFILE_FOOTPRINT) {
+                mgb_assert(std::get<1>(m_entry_stack.top()) == &entry);
+                MGB_LOCK_GUARD(m_lock);
+                m_entry_stack.pop();
+            }
+            m_profile.push_back(std::move(entry));
             return outputs;
         });
-        m_hooker_list.push_back(std::move(hooker));
+        if (flags & PROFILE_FOOTPRINT) {
+            hook_apply_on_var_node->apply_hook(
+                    [this](auto&& apply, const OpDef& def,
+                           VarNodeArray inputs) -> cg::OperatorNodeBase* {
+                        auto* operator_node = apply(def, std::move(inputs));
+                        std::remove_reference_t<decltype(m_entry_stack.top())>
+                                top;
+                        {
+                            MGB_LOCK_GUARD(m_lock);
+                            if (m_entry_stack.empty()) {
+                                return operator_node;
+                            }
+                            top = m_entry_stack.top();
+                        }
+                        auto [current_op, current_entry, thread_id] = top;
+                        if (current_op != &def ||
+                            thread_id != std::this_thread::get_id()) {
+                            return operator_node;
+                        }
+                        auto&& footprint_result =
+                                footprint.calc_footprint(operator_node);
+                        current_entry->memory = footprint_result.memory;
+                        current_entry->computation =
+                                footprint_result.computation;
+#if MGB_ENABLE_JSON
+                        current_entry->param = footprint_result.param;
+#endif
+                        return operator_node;
+                    });
+        }
+        m_hooker_list.push_back(std::move(hook_apply_on_physical_tensor));
+        m_hooker_list.push_back(std::move(hook_apply_on_var_node));
     });
 }
 
 void Profiler::stop() {
     m_hooker_list.clear();
-    for (auto& entry : *m_profile) {
+    for (auto& entry : m_profile) {
         entry.wait_device();
     }
 }
 
+void Profiler::clear() {
+    mgb_assert(m_entry_stack.empty(),
+               "entry_stack should be empty after profile");
+    mgb_assert(m_hooker_list.empty(), "hooks should be released");
+    m_profile.clear();
+    m_entry_count = 0;
+    m_device_timer.clear();
+    m_tensor_recorder.clear();
+}
+
 }  // namespace imperative
 
 }  // namespace mgb

imperative/src/include/megbrain/imperative/profiler.h

@@ -11,7 +11,10 @@
 
 #pragma once
 
-#include <variant>
+#include <any>
+#include <optional>
+#include <stack>
+#include <list>
 
 #include "megbrain/comp_node.h"
 #include "megbrain/graph/event.h"
@@ -19,27 +22,39 @@
 #include "megbrain/utils/timer.h"
 
 #include "megbrain/imperative/op_def.h"
-
-#include "megbrain/imperative/function_hook.h"
+#include "megbrain/imperative/physical_tensor.h"
 
 namespace mgb {
 namespace imperative {
 
-struct ProfileEntry{
+using ProfileTensor = std::tuple<size_t, std::vector<size_t>, CompNode>;
+
+struct ProfileEntry {
     using TimeClosure = std::function<double()>;
+    size_t id;
+    size_t parent;
     std::shared_ptr<OpDef> op;
+    //(host_begin, host_end)
     std::tuple<double, double> host;
+    //[(device, device_begin, device_end)]
     std::vector<std::tuple<CompNode, TimeClosure, TimeClosure>> device_list;
-    void wait_device(){
-        for(auto& [cn, begin, end]: device_list){
+    std::vector<ProfileTensor> inputs;
+    std::vector<ProfileTensor> outputs;
+    ssize_t memory = 0;
+    ssize_t computation = 0;
+#if MGB_ENABLE_JSON
+    std::shared_ptr<json::Value> param;
+#endif
+    void wait_device() {
+        for (auto& [cn, begin, end] : device_list) {
             MGB_MARK_USED_VAR(cn);
-            begin = [begin=begin()]{ return begin; };
-            end  = [end = end()]{ return end; };
+            begin = [begin = begin()] { return begin; };
+            end = [end = end()] { return end; };
         }
     }
 };
 
-using Profile = std::vector<ProfileEntry>;
+using Profile = std::list<ProfileEntry>;
 
 class DeviceTimer {
 public:
@@ -47,31 +62,54 @@ public:
     DeviceTimer() = default;
     void reset(thin_function<double()> host_timer);
     thin_function<double()> get_device_time(CompNode device);
+    void clear();
 
 private:
     CompNode::UnorderedMap<std::tuple<SharedEvent, double>> m_base_event_table;
+    thin_function<double()> m_host_timer;
+};
+
+class TensorRecorder {
+private:
+    // active tensors
+    std::unordered_map<Tensor*, std::tuple<std::weak_ptr<Tensor>, size_t>>
+            m_tensor_map;
+    size_t m_next_id;
+
+public:
+    size_t record_tensor(const TensorPtr& tensor);
+    void clear();
 };
 
 class Profiler {
 public:
-    Profiler(Profile* profile = nullptr) {
-        if (!profile) {
-            m_owned_profile = std::make_unique<Profile>();
-            profile = m_owned_profile.get();
-        }
-        m_profile = profile;
-    }
-    void start();
+    enum Flags {
+        PROFILE_FOOTPRINT = 1,
+    };
+
+public:
+    Profiler() = default;
+    // Start profiler by hook OpTrait
+    void start(uint32_t flags);
+    // Stop profiler and clean environment
     void stop();
-    Profile& get_profile() { return *m_profile; }
+    void clear();
+    Profile& get_profile();
 
 private:
     DeviceTimer m_device_timer;
     RealTimer m_host_timer;
-    Profile* m_profile;
+    Profile m_profile;
+    TensorRecorder m_tensor_recorder;
+    std::stack<std::tuple<const OpDef*, ProfileEntry*, std::thread::id>>
+            m_entry_stack;
+    // Hold profile owned by this Profiler
     std::unique_ptr<Profile> m_owned_profile;
-    std::vector<FunctionHooker<decltype(OpDef::apply_on_physical_tensor)>>
-            m_hooker_list;
+    // Hold hooks, cleared when stop
+    std::vector<std::any> m_hooker_list;
+    size_t m_entry_count = 0;
+    Spinlock m_lock;
+    std::unordered_map<Tensor*, std::weak_ptr<Tensor>> m_recorded_tensors;
 };
 
 }  // namespace imperative