Remove link copy queues

Author: adurang

offload/plugins-nextgen/level_zero/include/L0Device.h

@@ -84,18 +84,12 @@ class L0DeviceTLSTy {
   /// Main copy command list for each device
   ze_command_list_handle_t CopyCmdList = nullptr;
 
-  /// Link copy command list for each device
-  ze_command_list_handle_t LinkCopyCmdList = nullptr;
-
   /// Command queue for each device
   ze_command_queue_handle_t CmdQueue = nullptr;
 
   /// Main copy command queue for each device
   ze_command_queue_handle_t CopyCmdQueue = nullptr;
 
-  /// Link copy command queues for each device
-  ze_command_queue_handle_t LinkCopyCmdQueue = nullptr;
-
   /// Immediate command list for each device
   ze_command_list_handle_t ImmCmdList = nullptr;
 
@@ -106,20 +100,17 @@ class L0DeviceTLSTy {
   L0DeviceTLSTy() = default;
   ~L0DeviceTLSTy() {
     // assert all fields are nullptr on destruction
-    assert(!CmdList && !CopyCmdList && !LinkCopyCmdList && !CmdQueue &&
-           !CopyCmdQueue && !LinkCopyCmdQueue && !ImmCmdList &&
-           !ImmCopyCmdList &&
+    assert(!CmdList && !CopyCmdList && !CmdQueue && !CopyCmdQueue &&
+           !ImmCmdList && !ImmCopyCmdList &&
            "L0DeviceTLSTy destroyed without clearing resources");
   }
 
   L0DeviceTLSTy(const L0DeviceTLSTy &) = delete;
   L0DeviceTLSTy(L0DeviceTLSTy &&Other) {
     CmdList = std::exchange(Other.CmdList, nullptr);
     CopyCmdList = std::exchange(Other.CopyCmdList, nullptr);
-    LinkCopyCmdList = std::exchange(Other.LinkCopyCmdList, nullptr);
     CmdQueue = std::exchange(Other.CmdQueue, nullptr);
     CopyCmdQueue = std::exchange(Other.CopyCmdQueue, nullptr);
-    LinkCopyCmdQueue = std::exchange(Other.LinkCopyCmdQueue, nullptr);
     ImmCmdList = std::exchange(Other.ImmCmdList, nullptr);
     ImmCopyCmdList = std::exchange(Other.ImmCopyCmdList, nullptr);
   }
@@ -130,8 +121,6 @@ class L0DeviceTLSTy {
       CALL_ZE_EXIT_FAIL(zeCommandListDestroy, CmdList);
     if (CopyCmdList)
       CALL_ZE_EXIT_FAIL(zeCommandListDestroy, CopyCmdList);
-    if (LinkCopyCmdList)
-      CALL_ZE_EXIT_FAIL(zeCommandListDestroy, LinkCopyCmdList);
     if (ImmCmdList)
       CALL_ZE_EXIT_FAIL(zeCommandListDestroy, ImmCmdList);
     if (ImmCopyCmdList)
@@ -140,15 +129,11 @@ class L0DeviceTLSTy {
       CALL_ZE_EXIT_FAIL(zeCommandQueueDestroy, CmdQueue);
     if (CopyCmdQueue)
       CALL_ZE_EXIT_FAIL(zeCommandQueueDestroy, CopyCmdQueue);
-    if (LinkCopyCmdQueue)
-      CALL_ZE_EXIT_FAIL(zeCommandQueueDestroy, LinkCopyCmdQueue);
 
     CmdList = nullptr;
     CopyCmdList = nullptr;
-    LinkCopyCmdList = nullptr;
     CmdQueue = nullptr;
     CopyCmdQueue = nullptr;
-    LinkCopyCmdQueue = nullptr;
     ImmCmdList = nullptr;
     ImmCopyCmdList = nullptr;
   }
@@ -164,13 +149,6 @@ class L0DeviceTLSTy {
     CopyCmdList = _CopyCmdList;
   }
 
-  ze_command_list_handle_t getLinkCopyCmdList() const {
-    return LinkCopyCmdList;
-  }
-  void setLinkCopyCmdList(ze_command_list_handle_t _LinkCopyCmdList) {
-    LinkCopyCmdList = _LinkCopyCmdList;
-  }
-
   ze_command_list_handle_t getImmCmdList() const { return ImmCmdList; }
   void setImmCmdList(ze_command_list_handle_t ImmCmdListIn) {
     ImmCmdList = ImmCmdListIn;
@@ -190,13 +168,6 @@ class L0DeviceTLSTy {
   void setCopyCmdQueue(ze_command_queue_handle_t CopyCmdQueueIn) {
     CopyCmdQueue = CopyCmdQueueIn;
   }
-
-  ze_command_queue_handle_t getLinkCopyCmdQueue() const {
-    return LinkCopyCmdQueue;
-  }
-  void setLinkCopyCmdQueue(ze_command_queue_handle_t LinkCopyCmdQueueIn) {
-    LinkCopyCmdQueue = LinkCopyCmdQueueIn;
-  }
 };
 
 struct L0DeviceTLSTableTy
@@ -238,8 +209,6 @@ class L0DeviceTy final : public GenericDeviceTy {
   std::pair<uint32_t, uint32_t> ComputeOrdinal{UINT32_MAX, 0};
   /// Command queue group ordinals for copying
   std::pair<uint32_t, uint32_t> CopyOrdinal{UINT32_MAX, 0};
-  /// Command queue group ordinals and number of queues for link copy engines
-  std::pair<uint32_t, uint32_t> LinkCopyOrdinal{UINT32_MAX, 0};
 
   /// Command queue index for each device
   uint32_t ComputeIndex = 0;
@@ -416,10 +385,6 @@ class L0DeviceTy final : public GenericDeviceTy {
   bool hasMainCopyEngine() const { return CopyOrdinal.first != UINT32_MAX; }
   uint32_t getMainCopyEngine() const { return CopyOrdinal.first; }
 
-  uint32_t getLinkCopyEngine() const { return LinkCopyOrdinal.first; }
-  uint32_t getNumLinkCopyQueues() const { return LinkCopyOrdinal.second; }
-  bool hasLinkCopyEngine() const { return getNumLinkCopyQueues() > 0; }
-
   bool deviceRequiresImmCmdList() const {
     return isDeviceIPorNewer(0x05004000);
   }
@@ -474,8 +439,6 @@ class L0DeviceTy final : public GenericDeviceTy {
   Expected<ze_command_queue_handle_t> getCmdQueue();
   Expected<ze_command_list_handle_t> getCopyCmdList();
   Expected<ze_command_queue_handle_t> getCopyCmdQueue();
-  Expected<ze_command_list_handle_t> getLinkCopyCmdList();
-  Expected<ze_command_queue_handle_t> getLinkCopyCmdQueue();
   Expected<ze_command_list_handle_t> getImmCmdList();
   Expected<ze_command_list_handle_t> getImmCopyCmdList();
 

offload/plugins-nextgen/level_zero/src/L0Device.cpp

@@ -211,7 +211,6 @@ Error L0DeviceTy::initImpl(GenericPluginTy &Plugin) {
 
   CopyOrdinal = findCopyOrdinal();
 
-  LinkCopyOrdinal = findCopyOrdinal(true);
   IsAsyncEnabled =
       isDiscreteDevice() && Options.CommandMode != CommandModeTy::Sync;
   if (auto Err = MemAllocator.initDevicePools(*this, getPlugin().getOptions()))
@@ -945,8 +944,6 @@ L0DeviceTy::createImmCmdList(uint32_t Ordinal, uint32_t Index, bool InOrder) {
 /// Create an immediate command list for copying
 Expected<ze_command_list_handle_t> L0DeviceTy::createImmCopyCmdList() {
   uint32_t Ordinal = getMainCopyEngine();
-  if (Ordinal == UINT32_MAX)
-    Ordinal = getLinkCopyEngine();
   if (Ordinal == UINT32_MAX)
     Ordinal = getComputeEngine();
   return createImmCmdList(Ordinal, /*Index*/ 0);
@@ -980,9 +977,6 @@ Expected<ze_command_list_handle_t> L0DeviceTy::getCopyCmdList() {
     }
     return CmdList;
   }
-  // Use link copy engine if available
-  if (hasLinkCopyEngine())
-    return getLinkCopyCmdList();
   // Use compute engine otherwise
   return getCmdList();
 }
@@ -1002,58 +996,6 @@ Expected<ze_command_queue_handle_t> L0DeviceTy::getCopyCmdQueue() {
     }
     return CmdQueue;
   }
-  // Use link copy engine if available
-  if (hasLinkCopyEngine())
-    return getLinkCopyCmdQueue();
-  // Use compute engine otherwise
-  return getCmdQueue();
-}
-
-Expected<ze_command_list_handle_t> L0DeviceTy::getLinkCopyCmdList() {
-  // Use link copy engine if available
-  if (hasLinkCopyEngine()) {
-    auto &TLS = getTLS();
-    auto CmdList = TLS.getLinkCopyCmdList();
-    if (!CmdList) {
-      auto CmdListOrErr =
-          createCmdList(getZeContext(), getZeDevice(), getLinkCopyEngine(),
-                        ZE_COMMAND_LIST_FLAG_EXPLICIT_ONLY, getZeId());
-      if (!CmdListOrErr)
-        return CmdListOrErr.takeError();
-      CmdList = *CmdListOrErr;
-      TLS.setLinkCopyCmdList(CmdList);
-    }
-    return CmdList;
-  }
-  // Use main copy engine if available
-  if (hasMainCopyEngine())
-    return getCopyCmdList();
-  // Use compute engine otherwise
-  return getCmdList();
-}
-
-Expected<ze_command_queue_handle_t> L0DeviceTy::getLinkCopyCmdQueue() {
-  // Use link copy engine if available
-  if (hasLinkCopyEngine()) {
-    auto &TLS = getTLS();
-    auto CmdQueue = TLS.getLinkCopyCmdQueue();
-    if (!CmdQueue) {
-      // Try to use different copy engines for multiple threads
-      uint32_t Index =
-          __kmpc_global_thread_num(nullptr) % getNumLinkCopyQueues();
-      auto CmdQueueOrErr =
-          createCmdQueue(getZeContext(), getZeDevice(), getLinkCopyEngine(),
-                         Index, ZE_COMMAND_QUEUE_FLAG_EXPLICIT_ONLY, getZeId());
-      if (!CmdQueueOrErr)
-        return CmdQueueOrErr.takeError();
-      CmdQueue = *CmdQueueOrErr;
-      TLS.setLinkCopyCmdQueue(CmdQueue);
-    }
-    return CmdQueue;
-  }
-  // Use main copy engine if available
-  if (hasMainCopyEngine())
-    return getCopyCmdQueue();
   // Use compute engine otherwise
   return getCmdQueue();
 }