From 998f71a8c97a14c24700877a69f35c463977c17a Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Fri, 8 Jan 2021 11:54:53 +0800
Subject: [PATCH] perf(mge/imperative): do not busy wait in imperative's queue

GitOrigin-RevId: de8db5109679ed70e38fcecfbb99478aa9303b2b
---
 imperative/python/src/helper.h          | 6 ++++++
 imperative/src/impl/interpreter_impl.h  | 6 +++++-
 imperative/src/impl/physical_tensor.cpp | 5 ++++-
 3 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/imperative/python/src/helper.h b/imperative/python/src/helper.h
index 6a2fb1c93..b9f61e3d4 100644
--- a/imperative/python/src/helper.h
+++ b/imperative/python/src/helper.h
@@ -67,6 +67,12 @@ auto to_tuple(T begin, T end, pybind11::return_value_policy policy = pybind11::r
 class PyTaskDipatcher {
     struct Queue : mgb::AsyncQueueSC<std::function<void(void)>, Queue> {
         using Task = std::function<void(void)>;
+
+        // set max_spin=0 to prevent Queue fetch task in busy wait manner.
+        // this won't affect throughput when python interpreter is sending enough task,
+        // but will significantly save CPU time when waiting for task, e.g. wait for data input
+        Queue() : mgb::AsyncQueueSC<std::function<void(void)>, Queue>(0) {}
+
         void process_one_task(Task& f) {
             if (!Py_IsInitialized()) return;
             pybind11::gil_scoped_acquire _;
diff --git a/imperative/src/impl/interpreter_impl.h b/imperative/src/impl/interpreter_impl.h
index c1dd8447d..4a5e3fe9e 100644
--- a/imperative/src/impl/interpreter_impl.h
+++ b/imperative/src/impl/interpreter_impl.h
@@ -207,7 +207,11 @@ private:
     size_t m_enable_evict = 0;
 
     struct WorkQueue : AsyncQueueSC<Command, WorkQueue> {
-        WorkQueue(ChannelImpl* owner) : m_owner(owner) {
+        // set max_spin=0 to prevent Queue fetch task in busy wait manner.
+        // this won't affect throughput when python interpreter is sending enough task,
+        // but will significantly save CPU time when waiting for task, e.g. wait for data input
+        WorkQueue(ChannelImpl* owner)
+                : AsyncQueueSC<Command, WorkQueue>(0), m_owner(owner) {
             sys::set_thread_name("interpreter");
         }
         void process_one_task(Command& cmd) {
diff --git a/imperative/src/impl/physical_tensor.cpp b/imperative/src/impl/physical_tensor.cpp
index 19fa9a290..b4a76626d 100644
--- a/imperative/src/impl/physical_tensor.cpp
+++ b/imperative/src/impl/physical_tensor.cpp
@@ -30,7 +30,10 @@ class AsyncReleaser : public CompNodeDepedentObject {
         AsyncReleaser* m_par_releaser;
 
     public:
-        Waiter(AsyncReleaser* releaser) : m_par_releaser(releaser) {}
+        // disable busy wait by set max_spin=0 to save CPU cycle
+        Waiter(AsyncReleaser* releaser)
+                : AsyncQueueSC<WaiterParam, Waiter>(0),
+                  m_par_releaser(releaser) {}
 
         void process_one_task(WaiterParam& param) {
             if (param.event->finished()) {
-- 
GitLab