diff --git a/imperative/python/src/module_trace.h b/imperative/python/src/module_trace.h
index 8a9f9ce8..90706e98 100644
--- a/imperative/python/src/module_trace.h
+++ b/imperative/python/src/module_trace.h
@@ -14,7 +14,6 @@
 #include <list>
 #include "megbrain/imperative/transformations/trace.h"
 #include "megbrain/imperative/utils/map.h"
-#include "megbrain/imperative/utils/stats.h"
 
 #include "./tensor.h"
 
diff --git a/imperative/python/src/tensor.cpp b/imperative/python/src/tensor.cpp
index 0f219745..efbb0509 100644
--- a/imperative/python/src/tensor.cpp
+++ b/imperative/python/src/tensor.cpp
@@ -23,9 +23,9 @@
 #include "megbrain/imperative/transformations/symbol.h"
 #include "megbrain/imperative/transformations/trace.h"
 #include "megbrain/imperative/utils/map.h"
-#include "megbrain/imperative/utils/stats.h"
 #include "megbrain/opr/io.h"
 #include "megbrain/plugin/profiler.h"
+#include "megbrain/utils/stats.h"
 
 #include "./common.h"
 #include "./grad.h"
@@ -1367,9 +1367,9 @@ void init_tensor(py::module m) {
         return reprs;
     });
 
-    m.def("print_stats", [] { imperative::Stats::print(); });
+    m.def("print_stats", [] { Stats::print(); });
 
-    m.def("reset_stats", [] { imperative::Stats::reset(); });
+    m.def("reset_stats", [] { Stats::reset(); });
 
     m.def("_get_convert_inputs",
           []() -> bool { return DTypePromoteCfg::convert_input_enabled; });
diff --git a/imperative/python/src/tensor_utils.cpp b/imperative/python/src/tensor_utils.cpp
index a424eee0..1b6fd944 100644
--- a/imperative/python/src/tensor_utils.cpp
+++ b/imperative/python/src/tensor_utils.cpp
@@ -21,7 +21,6 @@
 #include "megbrain/imperative/transformations/symbol.h"
 #include "megbrain/imperative/transformations/trace.h"
 #include "megbrain/imperative/utils/map.h"
-#include "megbrain/imperative/utils/stats.h"
 #include "megbrain/opr/io.h"
 #include "megbrain/plugin/profiler.h"
 
diff --git a/imperative/src/impl/dispatch.cpp b/imperative/src/impl/dispatch.cpp
index 7f94d314..c982c6c2 100644
--- a/imperative/src/impl/dispatch.cpp
+++ b/imperative/src/impl/dispatch.cpp
@@ -14,7 +14,6 @@
 #include "megbrain/imperative/utils/debug.h"
 #include "megbrain/imperative/utils/helper.h"
 #include "megbrain/imperative/utils/map.h"
-#include "megbrain/imperative/utils/stats.h"
 
 namespace mgb {
 namespace imperative {
diff --git a/imperative/src/impl/interpreter/interpreter_impl.cpp b/imperative/src/impl/interpreter/interpreter_impl.cpp
index 6a5af98f..5b556430 100644
--- a/imperative/src/impl/interpreter/interpreter_impl.cpp
+++ b/imperative/src/impl/interpreter/interpreter_impl.cpp
@@ -19,7 +19,6 @@
 #include "megbrain/imperative/ops/backward_graph.h"
 #include "megbrain/imperative/ops/opr_attr.h"
 #include "megbrain/imperative/ops/utility.h"
-#include "megbrain/imperative/utils/stats.h"
 #include "megbrain/imperative/utils/to_string.h"
 
 #include "../blob_manager_impl.h"
diff --git a/imperative/src/impl/ops/elemwise.cpp b/imperative/src/impl/ops/elemwise.cpp
index 621e7f34..7d19b5e3 100644
--- a/imperative/src/impl/ops/elemwise.cpp
+++ b/imperative/src/impl/ops/elemwise.cpp
@@ -11,7 +11,6 @@
 
 #include "megbrain/imperative/opr_utility.h"
 #include "megbrain/imperative/ops/autogen.h"
-#include "megbrain/imperative/utils/stats.h"
 #include "megbrain/opr/basic_arith.h"
 #include "megbrain/opr/utility.h"
 
diff --git a/imperative/src/impl/ops/pooling.cpp b/imperative/src/impl/ops/pooling.cpp
index c7b4c59e..8281ac0f 100644
--- a/imperative/src/impl/ops/pooling.cpp
+++ b/imperative/src/impl/ops/pooling.cpp
@@ -11,7 +11,6 @@
 
 #include "megbrain/opr/dnn/pooling.h"
 #include "megbrain/imperative/ops/autogen.h"
-#include "megbrain/imperative/utils/stats.h"
 #include "megbrain/opr/utility.h"
 
 #include "megbrain/opr/internal/megdnn_opr_wrapper.h"
diff --git a/imperative/src/impl/transformation.cpp b/imperative/src/impl/transformation.cpp
index 32c0c4b3..2b3b326f 100644
--- a/imperative/src/impl/transformation.cpp
+++ b/imperative/src/impl/transformation.cpp
@@ -1,5 +1,4 @@
 #include "megbrain/imperative/transformation.h"
-#include "megbrain/imperative/utils/stats.h"
 
 namespace mgb {
 namespace imperative {
diff --git a/imperative/src/impl/transformations/eval.cpp b/imperative/src/impl/transformations/eval.cpp
index d0478921..5fda5437 100644
--- a/imperative/src/impl/transformations/eval.cpp
+++ b/imperative/src/impl/transformations/eval.cpp
@@ -10,7 +10,6 @@
  */
 
 #include "megbrain/imperative/transformations/eval.h"
-#include "megbrain/imperative/utils/stats.h"
 
 namespace mgb {
 namespace imperative {
diff --git a/imperative/src/impl/transformations/grad.cpp b/imperative/src/impl/transformations/grad.cpp
index b7660236..54e93100 100644
--- a/imperative/src/impl/transformations/grad.cpp
+++ b/imperative/src/impl/transformations/grad.cpp
@@ -15,7 +15,6 @@
 
 #include "megbrain/imperative/graph_cache.h"
 #include "megbrain/imperative/resource_manager.h"
-#include "megbrain/imperative/utils/stats.h"
 
 #include <range/v3/all.hpp>
 
diff --git a/imperative/src/impl/transformations/scalar.cpp b/imperative/src/impl/transformations/scalar.cpp
index 89ee1f46..2d6fd7c9 100644
--- a/imperative/src/impl/transformations/scalar.cpp
+++ b/imperative/src/impl/transformations/scalar.cpp
@@ -13,7 +13,6 @@
 
 #include "megbrain/imperative/ops/autogen.h"
 #include "megbrain/imperative/ops/utility.h"
-#include "megbrain/imperative/utils/stats.h"
 
 namespace mgb {
 namespace imperative {
diff --git a/imperative/src/include/megbrain/imperative/utils/stats.h b/imperative/src/include/megbrain/imperative/utils/stats.h
deleted file mode 100644
index 48d36171..00000000
--- a/imperative/src/include/megbrain/imperative/utils/stats.h
+++ /dev/null
@@ -1,199 +0,0 @@
-#pragma once
-
-#include <chrono>
-#include <iostream>
-#include <map>
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace mgb {
-namespace imperative {
-namespace stats {
-
-#define MGE_ENABLE_STATS 0
-
-class Timer {
-public:
-    using clock_t = std::chrono::system_clock;
-
-private:
-    clock_t::duration m_duration = clock_t::duration{0};
-    size_t m_timing = 0;
-    std::string m_name;
-    uint64_t m_count = 0;
-    size_t m_enabled = 1;
-    bool m_default_enabled = true;
-
-    struct TimeScopeRecursive {
-        Timer& timer;
-        clock_t::time_point start;
-        bool released = false;
-
-        TimeScopeRecursive(Timer& timer) : timer(timer) {
-            if (timer.m_enabled && !timer.m_timing++) {
-                start = clock_t::now();
-            }
-        }
-
-        ~TimeScopeRecursive() { release(); }
-
-        void release() {
-            if (released) {
-                return;
-            }
-            if (timer.m_enabled) {
-                if (!--timer.m_timing) {
-                    auto duration = (clock_t::now() - start);
-                    timer.m_duration += duration;
-                }
-                timer.m_count++;
-            }
-            released = true;
-        }
-    };
-
-    struct EnableScope {
-        Timer& timer;
-        bool released = false;
-
-        EnableScope(Timer& timer) : timer(timer) { timer.m_enabled++; }
-
-        ~EnableScope() { release(); }
-
-        void release() {
-            if (released) {
-                return;
-            }
-            timer.m_enabled--;
-            released = true;
-        }
-    };
-
-public:
-    Timer(std::string name, bool default_enabled = true);
-
-    std::string name() { return m_name; }
-    auto time_scope_recursive() { return TimeScopeRecursive(*this); };
-    auto enable_scope() { return EnableScope(*this); }
-    void reset() {
-        m_duration = clock_t::duration{0};
-        m_count = 0;
-        m_enabled = m_default_enabled ? 1 : 0;
-    }
-
-    clock_t::duration get() const { return m_duration; }
-    uint64_t count() const { return m_count; }
-};
-}  // namespace stats
-
-struct Stats {
-    struct TimerNode {
-        std::map<std::string, std::unique_ptr<TimerNode>> children;
-        stats::Timer* timer = nullptr;
-
-        TimerNode() {}
-    };
-
-    static inline TimerNode sm_root;
-
-    // register your timers here
-    // for example:
-    //
-    // static inline stats::Timer mytimer;
-    //
-    // then use MGE_TIMER_SCOPE(mytimer) to collect durations in your code
-
-    static std::pair<long, long> print_node(
-            std::string name, TimerNode& node, size_t indent = 0) {
-        auto print_indent = [&] {
-            for (size_t i = 0; i < indent; ++i) {
-                printf(" ");
-            }
-        };
-        long ns = 0, count = 0;
-        if (auto* timer = node.timer) {
-            print_indent();
-            printf("%s costs %'ld ns, hits %'ld times\n", name.c_str(),
-                   (long)timer->get().count(), (long)timer->count());
-            ns = timer->get().count();
-            count = timer->count();
-        }
-        if (!node.children.empty()) {
-            bool collect_children = node.timer == nullptr;
-            if (collect_children) {
-                print_indent();
-                printf("%s:\n", name.c_str());
-            }
-            long ns = 0, count = 0;
-            for (auto&& child : node.children) {
-                auto [child_ns, child_count] =
-                        print_node(child.first, *child.second, indent + 4);
-                if (collect_children) {
-                    ns += child_ns;
-                    count += child_count;
-                }
-            }
-            if (collect_children) {
-                print_indent();
-                printf("total costs %'ld ns, hits %'ld times\n", ns, count);
-            }
-        }
-        return {ns, count};
-    }
-
-    static void print() {
-        for (auto&& child : sm_root.children) {
-            print_node(child.first, *child.second);
-        }
-    }
-
-    static void reset() {
-        auto reset_node = [](TimerNode& node, auto&& reset_node) -> void {
-            if (auto* timer = node.timer) {
-                timer->reset();
-            }
-            for (auto&& child : node.children) {
-                reset_node(*child.second, reset_node);
-            }
-        };
-        reset_node(sm_root, reset_node);
-    }
-};
-
-inline stats::Timer::Timer(std::string name, bool default_enabled)
-        : m_name(name), m_default_enabled(default_enabled) {
-    std::vector<std::string> terms;
-    Stats::TimerNode* node = &Stats::sm_root;
-    while (true) {
-        auto pos = name.find(".");
-        if (pos == std::string::npos) {
-            auto& child = node->children[name];
-            child = std::make_unique<Stats::TimerNode>();
-            node = child.get();
-            node->timer = this;
-            break;
-        } else {
-            auto& child = node->children[name.substr(0, pos)];
-            if (!child) {
-                child = std::make_unique<Stats::TimerNode>();
-            }
-            node = child.get();
-            name = name.substr(pos + 1);
-        }
-    }
-}
-
-#if MGE_ENABLE_STATS
-#define MGE_TIMER_SCOPE(name)         auto name = Stats::name.time_scope_recursive()
-#define MGE_TIMER_SCOPE_RELEASE(name) name.release()
-#define MGE_TIMER_SCOPE_ENABLE(name)  auto name = Stats::name.enable_scope()
-#else
-#define MGE_TIMER_SCOPE(name)         (void)0
-#define MGE_TIMER_SCOPE_RELEASE(name) (void)0
-#define MGE_TIMER_SCOPE_ENABLE(name)  (void)0
-#endif
-
-}  // namespace imperative
-}  // namespace mgb
diff --git a/imperative/src/include/megbrain/imperative/value.h b/imperative/src/include/megbrain/imperative/value.h
index 21541443..c877d37f 100644
--- a/imperative/src/include/megbrain/imperative/value.h
+++ b/imperative/src/include/megbrain/imperative/value.h
@@ -23,7 +23,6 @@
 #include "megbrain/imperative/utils/debug.h"
 #include "megbrain/imperative/utils/local_ptr.h"
 #include "megbrain/imperative/utils/span.h"
-#include "megbrain/imperative/utils/stats.h"
 
 namespace mgb {
 namespace imperative {
diff --git a/src/core/impl/utils/stats.cpp b/src/core/impl/utils/stats.cpp
new file mode 100644
index 00000000..93c5ffd4
--- /dev/null
+++ b/src/core/impl/utils/stats.cpp
@@ -0,0 +1,89 @@
+#include "megbrain/utils/stats.h"
+
+namespace mgb {
+
+Stats::TimerNode Stats::sm_root;
+
+stats::Timer& Stats::get_timer(std::string name) {
+    auto full_name = name;
+    Stats::TimerNode* node = &Stats::sm_root;
+    while (true) {
+        auto pos = name.find("_");
+        if (pos == std::string::npos) {
+            auto& child = node->children[name];
+            child = std::make_unique<Stats::TimerNode>();
+            node = child.get();
+            auto& timer = node->timer;
+            if (!timer) {
+                timer = std::make_unique<stats::Timer>(full_name);
+            }
+            return *timer;
+        } else {
+            auto& child = node->children[name.substr(0, pos)];
+            if (!child) {
+                child = std::make_unique<Stats::TimerNode>();
+            }
+            node = child.get();
+            name = name.substr(pos + 1);
+        }
+    }
+}
+
+std::pair<long, long> Stats::print_node(
+        std::string name, TimerNode& node, size_t indent) {
+    auto print_indent = [&] {
+        for (size_t i = 0; i < indent; ++i) {
+            printf(" ");
+        }
+    };
+    long ns = 0, count = 0;
+    if (auto& timer = node.timer) {
+        print_indent();
+        printf("%s costs %'ld ns, hits %'ld times\n", name.c_str(),
+               (long)timer->get().count(), (long)timer->count());
+        ns = timer->get().count();
+        count = timer->count();
+    }
+    if (!node.children.empty()) {
+        bool collect_children = node.timer == nullptr;
+        if (collect_children) {
+            print_indent();
+            printf("%s:\n", name.c_str());
+        }
+        long ns = 0, count = 0;
+        for (auto&& child : node.children) {
+            auto&& child_res = print_node(child.first, *child.second, indent + 4);
+            auto&& child_ns = child_res.first;
+            auto&& child_count = child_res.second;
+            if (collect_children) {
+                ns += child_ns;
+                count += child_count;
+            }
+        }
+        if (collect_children) {
+            print_indent();
+            printf("total costs %'ld ns, hits %'ld times\n", ns, count);
+        }
+    }
+    return {ns, count};
+}
+
+void Stats::print() {
+    for (auto&& child : sm_root.children) {
+        print_node(child.first, *child.second);
+    }
+}
+
+void Stats::reset() {
+    auto reset_node = [](TimerNode& node, auto&& reset_node) -> void {
+        if (auto& timer = node.timer) {
+            timer->reset();
+        }
+        for (auto&& child : node.children) {
+            reset_node(*child.second, reset_node);
+        }
+    };
+    reset_node(sm_root, reset_node);
+}
+
+}  // namespace mgb
\ No newline at end of file
diff --git a/src/core/include/megbrain/utils/stats.h b/src/core/include/megbrain/utils/stats.h
new file mode 100644
index 00000000..97ba80c4
--- /dev/null
+++ b/src/core/include/megbrain/utils/stats.h
@@ -0,0 +1,135 @@
+#pragma once
+
+#include <chrono>
+#include <iostream>
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "megbrain/common.h"
+
+namespace mgb {
+namespace stats {
+
+#define MGE_ENABLE_STATS 1
+
+class Timer {
+public:
+    using clock_t = std::chrono::system_clock;
+
+private:
+    clock_t::duration m_duration = clock_t::duration{0};
+    size_t m_timing = 0;
+    std::string m_name;
+    uint64_t m_count = 0;
+    size_t m_enabled = 1;
+    bool m_default_enabled = true;
+
+    struct TimeScopeRecursive {
+        Timer& timer;
+        clock_t::time_point start;
+        bool released = false;
+
+        TimeScopeRecursive(Timer& timer) : timer(timer) {
+            if (timer.m_enabled && !timer.m_timing++) {
+                start = clock_t::now();
+            }
+        }
+
+        ~TimeScopeRecursive() { release(); }
+
+        void release() {
+            if (released) {
+                return;
+            }
+            if (timer.m_enabled) {
+                if (!--timer.m_timing) {
+                    auto duration = (clock_t::now() - start);
+                    timer.m_duration += duration;
+                }
+                timer.m_count++;
+            }
+            released = true;
+        }
+    };
+
+    struct EnableScope {
+        Timer& timer;
+        bool released = false;
+
+        EnableScope(Timer& timer) : timer(timer) { timer.m_enabled++; }
+
+        ~EnableScope() { release(); }
+
+        void release() {
+            if (released) {
+                return;
+            }
+            timer.m_enabled--;
+            released = true;
+        }
+    };
+
+public:
+    Timer(std::string name, bool default_enabled = true)
+            : m_name(name), m_default_enabled(default_enabled){};
+
+    std::string name() { return m_name; }
+    auto time_scope_recursive() { return TimeScopeRecursive(*this); };
+    auto enable_scope() { return EnableScope(*this); }
+    void reset() {
+        m_duration = clock_t::duration{0};
+        m_count = 0;
+        m_enabled = m_default_enabled ? 1 : 0;
+    }
+
+    clock_t::duration get() const { return m_duration; }
+    uint64_t count() const { return m_count; }
+};
+}  // namespace stats
+
+struct Stats {
+private:
+    struct TimerNode {
+        std::map<std::string, std::unique_ptr<TimerNode>> children;
+        std::unique_ptr<stats::Timer> timer;
+    };
+
+    static TimerNode sm_root;
+
+    // don't register your timers here
+    // use MGE_TIMER_SCOPE(mytimer) to collect durations in your code
+public:
+    MGE_WIN_DECLSPEC_FUC static stats::Timer& get_timer(std::string name);
+
+    MGE_WIN_DECLSPEC_FUC static std::pair<long, long> print_node(
+            std::string name, TimerNode& node, size_t indent = 0);
+
+    MGE_WIN_DECLSPEC_FUC static void print();
+
+    MGE_WIN_DECLSPEC_FUC static void reset();
+};
+
+#if MGE_ENABLE_STATS
+
+#define MGE_TIMER_SCOPE(name)                                  \
+    static auto& _timer_##name = mgb::Stats::get_timer(#name); \
+    auto name = _timer_##name.time_scope_recursive()
+
+#define MGE_TIMER_SCOPE_RELEASE(name) name.release()
+
+#define MGE_TIMER_SCOPE_ENABLE(name)                           \
+    static auto& _timer_##name = mgb::Stats::get_timer(#name); \
+    auto name = _timer_##name.enable_scope()
+
+#else
+
+#define MGE_TIMER_SCOPE(name)         (void)0
+#define MGE_TIMER_SCOPE_RELEASE(name) (void)0
+#define MGE_TIMER_SCOPE_ENABLE(name)  (void)0
+
+#endif
+
+}  // namespace mgb