From 89aafbc3dbb712559f610454c5de0546834d64b9 Mon Sep 17 00:00:00 2001
From: Paulius Velesko <pvelesko@pglc.io>
Date: Wed, 4 Sep 2024 07:09:28 +0300
Subject: [PATCH] remove dubious locks

---
 CMakeLists.txt                          |  1 -
 chipStarConfig.hh.in                    |  2 --
 src/CHIPBackend.hh                      |  5 -----
 src/backend/Level0/CHIPBackendLevel0.cc |  9 ---------
 src/backend/OpenCL/CHIPBackendOpenCL.cc | 14 --------------
 5 files changed, 31 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index a3db1c018..6c40a19c6 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -212,7 +212,6 @@ option(CHIP_SET_RPATH "Add CMAKE_INSTALL_PREFIX/lib to the RPATH for chipStar ex
 option(CHIP_ENABLE_UNCOMPILABLE_TESTS "Enable tests which are known to not compile" OFF)
 option(CHIP_BUILD_TESTS "Enable build_tests target" ON)
 option(CHIP_BUILD_SAMPLES "Build samples" ON)
-option(CHIP_DUBIOUS_LOCKS "Enable locks that don't seem necessary but make a lot of valgrind issues go away" OFF)
 option(CHIP_USE_EXTERNAL_HIP_TESTS "Use Catch2 tests from the hip-tests submodule" OFF)
 option(CHIP_ENABLE_NON_COMPLIANT_DEVICELIB_CODE "Enable non-compliant devicelib code such as calling LLVM builtins from inside kernel code. Enables certain unsigned long devicelib func variants" OFF)
 option(CHIP_FAST_MATH "Use native_ OpenCL functions which are fast but their precision is implementation defined" OFF)
diff --git a/chipStarConfig.hh.in b/chipStarConfig.hh.in
index 542010324..f31e36cc8 100644
--- a/chipStarConfig.hh.in
+++ b/chipStarConfig.hh.in
@@ -51,8 +51,6 @@
 
 #cmakedefine CHIP_USE_INTEL_USM
 
-#cmakedefine CHIP_DUBIOUS_LOCKS
-
 #cmakedefine CHIP_L0_FIRST_TOUCH
 
 #cmakedefine CHIP_ENABLE_NON_COMPLIANT_DEVICELIB_CODE
diff --git a/src/CHIPBackend.hh b/src/CHIPBackend.hh
index 1ed2e6749..ae5ab5966 100644
--- a/src/CHIPBackend.hh
+++ b/src/CHIPBackend.hh
@@ -1857,11 +1857,6 @@ public:
    */
   virtual void trackEvent(const std::shared_ptr<chipstar::Event> &Event);
 
-#ifdef CHIP_DUBIOUS_LOCKS
-  std::mutex DubiousLockOpenCL;
-  std::mutex DubiousLockLevel0;
-#endif
-
   virtual chipstar::ExecItem *createExecItem(dim3 GirdDim, dim3 BlockDim,
                                              size_t SharedMem,
                                              hipStream_t ChipQueue) = 0;
diff --git a/src/backend/Level0/CHIPBackendLevel0.cc b/src/backend/Level0/CHIPBackendLevel0.cc
index 527c2bef2..972501bcf 100644
--- a/src/backend/Level0/CHIPBackendLevel0.cc
+++ b/src/backend/Level0/CHIPBackendLevel0.cc
@@ -808,9 +808,6 @@ CHIPQueueLevel0::~CHIPQueueLevel0() {
   // The application must not call this function from
   // simultaneous threads with the same command queue handle.
   // Done. Destructor should not be called by multiple threads
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockLevel0)
-#endif
   if (zeCmdQOwnership_) {
     zeStatus = zeCommandQueueDestroy(ZeCmdQ_);
   } else {
@@ -972,9 +969,6 @@ CHIPQueueLevel0::CHIPQueueLevel0(CHIPDeviceLevel0 *ChipDev,
   ZeDev_ = ChipDevLz_->get();
 
   logTrace("CHIPQueueLevel0 constructor called via Flags and Priority");
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockLevel0)
-#endif
   zeStatus = zeCommandQueueCreate(ZeCtx_, ZeDev_, &QueueDescriptor_, &ZeCmdQ_);
   CHIPERR_CHECK_LOG_AND_THROW_TABLE(zeCommandQueueCreate);
 
@@ -1462,9 +1456,6 @@ CHIPQueueLevel0::memCopyAsyncImpl(void *Dst, const void *Src, size_t Size,
 }
 
 void CHIPQueueLevel0::finish() {
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockLevel0)
-#endif
   auto LastEvent = getLastEvent();
   if (LastEvent)
     LastEvent->wait();
diff --git a/src/backend/OpenCL/CHIPBackendOpenCL.cc b/src/backend/OpenCL/CHIPBackendOpenCL.cc
index 17aa8fa29..79ed6b163 100644
--- a/src/backend/OpenCL/CHIPBackendOpenCL.cc
+++ b/src/backend/OpenCL/CHIPBackendOpenCL.cc
@@ -1285,12 +1285,7 @@ CHIPQueueOpenCL::launchImpl(chipstar::ExecItem *ExecItem) {
   const size_t Local[NumDims] = {BlockDim.x, BlockDim.y, BlockDim.z};
 
   logTrace("Launch GLOBAL: {} {} {}", Global[0], Global[1], Global[2]);
-
   logTrace("Launch LOCAL: {} {} {}", Local[0], Local[1], Local[2]);
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockOpenCL);
-#endif
-
   auto AllocationsToKeepAlive = annotateIndirectPointers(
       *OclContext, Kernel->getModule()->getInfo(), KernelHandle);
 
@@ -1431,9 +1426,6 @@ CHIPQueueOpenCL::memCopyAsyncImpl(void *Dst, const void *Src, size_t Size,
         std::static_pointer_cast<CHIPEventOpenCL>(Event)->getNativePtr());
     CHIPERR_CHECK_LOG_AND_THROW_TABLE(clEnqueueMarker);
   } else {
-#ifdef CHIP_DUBIOUS_LOCKS
-    LOCK(Backend->DubiousLockOpenCL)
-#endif
     auto [EventsToWait, EventLocks] = getSyncQueuesLastEvents(Event, false);
     std::vector<cl_event> SyncQueuesEventHandles =
         getOpenCLHandles(EventsToWait);
@@ -1518,9 +1510,6 @@ CHIPQueueOpenCL::memCopyAsyncImpl(void *Dst, const void *Src, size_t Size,
 }
 
 void CHIPQueueOpenCL::finish() {
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockOpenCL)
-#endif
   clStatus = get()->finish();
   CHIPERR_CHECK_LOG_AND_THROW_TABLE(clFinish);
   this->LastEvent_ = nullptr;
@@ -1627,9 +1616,6 @@ CHIPQueueOpenCL::memPrefetchImpl(const void *Ptr, size_t Count) {
 
 std::shared_ptr<chipstar::Event> CHIPQueueOpenCL::enqueueBarrierImpl(
     const std::vector<std::shared_ptr<chipstar::Event>> &EventsToWaitFor) {
-#ifdef CHIP_DUBIOUS_LOCKS
-  LOCK(Backend->DubiousLockOpenCL)
-#endif
   std::shared_ptr<chipstar::Event> Event =
       static_cast<CHIPBackendOpenCL *>(Backend)->createEventShared(
           this->ChipContext_);