From d6b692f4aac58184d60b157906abef43c78d81da Mon Sep 17 00:00:00 2001
From: xu-hong26 <xuhong26@hisilicon.com>
Date: Mon, 14 Aug 2023 19:45:18 +0800
Subject: [PATCH] TaskQueue optimize

---
 torch_npu/csrc/core/npu/NPUQueue.cpp | 40 +++++++++++++++-------------
 torch_npu/csrc/core/npu/NPUQueue.h   |  2 ++
 2 files changed, 23 insertions(+), 19 deletions(-)

diff --git a/torch_npu/csrc/core/npu/NPUQueue.cpp b/torch_npu/csrc/core/npu/NPUQueue.cpp
index 632e24d779e..0f7a68fb4c6 100644
--- a/torch_npu/csrc/core/npu/NPUQueue.cpp
+++ b/torch_npu/csrc/core/npu/NPUQueue.cpp
@@ -296,17 +296,6 @@ void Repository::Enqueue(void* cur_paras) {
       continue;
     }
     __sync_synchronize();
-    while (!IsReadWorking()) {
-      s = eventfd_write(efd_read, u);
-      if (s != 0) {
-        if (errno == EINTR) {
-          continue;
-        }
-        ASCEND_LOGE("notify consumer failed!! s=%zd, errno=%s", s, strerror(errno));
-        return;
-      }
-      break;
-    }
   }
   SetWriteWorking(false);
 }
@@ -334,14 +323,7 @@ void Repository::Dequeue() {
       SetReadWorking(false);
       __sync_synchronize();
       if (IsEmptyQueue()) {
-        s = eventfd_read(efd_read, &u);
-        if (s != 0) {
-          if (errno == EINTR) {
-            continue;
-          }
-          ASCEND_LOGE("waiting enqueue failed. s=%zd, errno=%s.", s, strerror(errno));
-          return;
-        }
+        WaitForTaskArrival();
         SetReadWorking(true);
       }
       continue;
@@ -376,6 +358,26 @@ void Repository::Dequeue() {
   SetReadWorking(false);
 }
 
+void Repository::WaitForTaskArrival() {
+  uint64_t loop_cnt = 0;
+  struct timeval delay = {0, 0};
+
+  while (IsEmptyQueue()) {
+    if (GetStatus() != RepoStatus::RUN) {
+      break;
+    }
+    loop_cnt++;
+    if (loop_cnt < TASK_ARRIVAL_LOOP_CNT) {
+      continue;
+    }
+    delay.tv_usec = 1;
+    select(0, nullptr, nullptr, nullptr, &delay);
+    __sync_synchronize();
+  }
+
+  return;
+}
+
 void Repository::ReleaseResource() {
   manager().DeInit(datas);
   if (efd_read > 0) {
diff --git a/torch_npu/csrc/core/npu/NPUQueue.h b/torch_npu/csrc/core/npu/NPUQueue.h
index 7295fe04487..8ce46d87a88 100644
--- a/torch_npu/csrc/core/npu/NPUQueue.h
+++ b/torch_npu/csrc/core/npu/NPUQueue.h
@@ -11,6 +11,7 @@
 
 namespace c10_npu {
 
+#define TASK_ARRIVAL_LOOP_CNT (3030000) // about 10ms
 struct sring_idx {
   bool working = false;
   volatile unsigned int idx = 0;
@@ -96,6 +97,7 @@ private:
   bool IsReadWorking() const {return read_idx.working;};
   bool WriteQueue(void* cur_paras);
   bool ReadQueue();
+  void WaitForTaskArrival();
 
 private:
   void* datas = nullptr;
-- 
Gitee