From 91f4cdcfc28543a167ef98f23db47840c2342f18 Mon Sep 17 00:00:00 2001
From: Paulius Velesko <pvelesko@pglc.io>
Date: Sat, 12 Oct 2024 16:16:14 +0300
Subject: [PATCH] Queue::finish fence fix

* event record uses regular command list enqueue. Finishing the queue in this case needs to wait not only on last even but on this fence.
---
 src/backend/Level0/CHIPBackendLevel0.cc    | 26 +++++++++++++++++++++-
 src/backend/Level0/CHIPBackendLevel0.hh    |  8 +++++++
 src/backend/Level0/zeHipErrorConversion.hh |  7 ++++++
 3 files changed, 40 insertions(+), 1 deletion(-)

diff --git a/src/backend/Level0/CHIPBackendLevel0.cc b/src/backend/Level0/CHIPBackendLevel0.cc
index 986b16cca..8e8448fe9 100644
--- a/src/backend/Level0/CHIPBackendLevel0.cc
+++ b/src/backend/Level0/CHIPBackendLevel0.cc
@@ -1472,6 +1472,25 @@ void CHIPQueueLevel0::finish() {
   if (LastEvent)
     LastEvent->wait();
 
+  if (ZeFenceLast_) {
+    auto BackendLz = static_cast<CHIPBackendLevel0 *>(Backend);
+    LOCK(BackendLz->EventsMtx);
+    LOCK(BackendLz->ActiveCmdListsMtx);
+    auto it = std::find_if(BackendLz->ActiveCmdLists.begin(),
+                           BackendLz->ActiveCmdLists.end(),
+                           [this](const auto &CmdList) {
+                             return CmdList->getFence() == ZeFenceLast_;
+                           });
+
+    if (it != BackendLz->ActiveCmdLists.end()) {
+      (*it)->wait();
+      BackendLz->ActiveCmdLists.erase(it);
+      return;
+    }
+
+    ZeFenceLast_ = nullptr;
+  }
+
   if (zeCmdQOwnership_) {
     zeStatus = zeCommandQueueSynchronize(ZeCmdQ_,
                                          ChipEnvVars.getL0EventTimeout() * 1e9);
@@ -1486,7 +1505,11 @@ void CHIPQueueLevel0::finish() {
 void CHIPQueueLevel0::executeCommandList(
     Borrowed<FencedCmdList> &CmdList, std::shared_ptr<chipstar::Event> Event) {
   updateLastEvent(Event);
-  CmdList->execute(getCmdQueue());
+
+  CmdList->execute(getCmdQueue()); // creates fence
+  ZeFenceLast_ = CmdList->getFence();
+  assert(ZeFenceLast_ && "Fence pointer is null");
+
   auto BackendLz = static_cast<CHIPBackendLevel0 *>(Backend);
   LOCK(BackendLz->ActiveCmdListsMtx);
   BackendLz->ActiveCmdLists.push_back(std::move(CmdList));
@@ -1624,6 +1647,7 @@ void CHIPBackendLevel0::uninitialize() {
     EventMonitor_->Stop = true;
   }
   EventMonitor_->join();
+  ActiveCmdLists.clear();
   return;
 }
 
diff --git a/src/backend/Level0/CHIPBackendLevel0.hh b/src/backend/Level0/CHIPBackendLevel0.hh
index 899afff37..61a5796f8 100644
--- a/src/backend/Level0/CHIPBackendLevel0.hh
+++ b/src/backend/Level0/CHIPBackendLevel0.hh
@@ -213,7 +213,14 @@ public:
     CHIPERR_CHECK_LOG_AND_ABORT("Failed to create command list");
   }
 
+  void wait() {
+    zeStatus = zeFenceHostSynchronize(ZeFence_, UINT64_MAX);
+    CHIPERR_CHECK_LOG_AND_THROW_TABLE(zeFenceHostSynchronize);
+  }
+
   bool reset() {
+    zeStatus = zeFenceReset(ZeFence_);
+    CHIPERR_CHECK_LOG_AND_ABORT("Failed to reset fence");
     zeStatus = zeCommandListReset(ZeCmdList_);
     CHIPERR_CHECK_LOG_AND_ABORT("Failed to reset command list");
     return true;
@@ -253,6 +260,7 @@ protected:
   ze_device_handle_t ZeDev_;
   CHIPDeviceLevel0 *ChipDevLz_;
   CHIPContextLevel0 *ChipCtxLz_;
+  ze_fence_handle_t ZeFenceLast_ = nullptr;
 
   // The shared memory buffer
   void *SharedBuf_;
diff --git a/src/backend/Level0/zeHipErrorConversion.hh b/src/backend/Level0/zeHipErrorConversion.hh
index 0b0ba6fe5..0fa2ff168 100644
--- a/src/backend/Level0/zeHipErrorConversion.hh
+++ b/src/backend/Level0/zeHipErrorConversion.hh
@@ -318,6 +318,13 @@ const std::unordered_map<void *, ze_hip_error_map_t> ZE_HIP_ERROR_MAPS = {
       {ZE_RESULT_ERROR_DEVICE_LOST, hipErrorNotInitialized},
       {ZE_RESULT_ERROR_INVALID_NULL_HANDLE, hipErrorInvalidResourceHandle},
       {ZE_RESULT_NOT_READY, hipErrorNotReady}}},
+    {(void *)&zeFenceHostSynchronize,
+     {{ZE_RESULT_SUCCESS, hipSuccess},
+      {ZE_RESULT_ERROR_UNINITIALIZED, hipErrorNotInitialized},
+      {ZE_RESULT_ERROR_DEVICE_LOST, hipErrorNotInitialized},
+      {ZE_RESULT_ERROR_INVALID_NULL_HANDLE, hipErrorInvalidResourceHandle},
+      {ZE_RESULT_ERROR_INVALID_SYNCHRONIZATION_OBJECT, hipErrorInvalidValue},
+      {ZE_RESULT_NOT_READY, hipErrorNotReady}}},
     {(void *)&zeCommandListAppendMemoryFill,
      {{ZE_RESULT_SUCCESS, hipSuccess},
       {ZE_RESULT_ERROR_UNINITIALIZED, hipErrorNotInitialized},