address review comments

Author: adurang

offload/plugins-nextgen/level_zero/include/L0Memory.h

@@ -200,8 +200,6 @@ class MemAllocatorTy {
     std::vector<std::pair<uint64_t, uint64_t>> BucketStats;
     /// Need to zero-initialize after L0 allocation
     bool ZeroInit = false;
-    /// Zero-initialized values to be copied to device
-    std::vector<char> ZeroInitValue;
 
     /// Get bucket ID from the specified allocation size.
     uint32_t getBucketId(size_t Size) {
@@ -394,6 +392,10 @@ class MemAllocatorTy {
 
   /// Perform copy operation
   int32_t enqueueMemCopy(void *Dst, const void *Src, size_t Size);
+
+  /// Perform memory fill operation
+  int32_t enqueueMemSet(void *Dst, int8_t Value, size_t Size);
+
 }; /// MemAllocatorTy
 
 // simple generic wrapper to reuse objects

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

@@ -701,14 +701,12 @@ int32_t L0DeviceTy::enqueueMemCopy(void *Dst, const void *Src, size_t Size,
                                    bool UseCopyEngine) {
   ze_command_list_handle_t CmdList = nullptr;
   ze_command_queue_handle_t CmdQueue = nullptr;
-  ze_event_handle_t Event = nullptr;
 
   if (useImmForCopy()) {
     CmdList = UseCopyEngine ? getImmCopyCmdList() : getImmCmdList();
-    Event = getEvent();
     CALL_ZE_RET_FAIL(zeCommandListAppendMemoryCopy, CmdList, Dst, Src, Size,
-                     Event, 0, nullptr);
-    CALL_ZE_RET_FAIL(zeEventHostSynchronize, Event, UINT64_MAX);
+                     nullptr, 0, nullptr);
+    CALL_ZE_RET_FAIL(zeCommandListHostSynchronize, CmdList, UINT64_MAX);
   } else {
     if (UseCopyEngine) {
       CmdList = getCopyCmdList();
@@ -719,7 +717,7 @@ int32_t L0DeviceTy::enqueueMemCopy(void *Dst, const void *Src, size_t Size,
     }
 
     CALL_ZE_RET_FAIL(zeCommandListAppendMemoryCopy, CmdList, Dst, Src, Size,
-                     Event, 0, nullptr);
+                     nullptr, 0, nullptr);
     CALL_ZE_RET_FAIL(zeCommandListClose, CmdList);
     CALL_ZE_RET_FAIL_MTX(zeCommandQueueExecuteCommandLists, getMutex(),
                          CmdQueue, 1, &CmdList, nullptr);

offload/plugins-nextgen/level_zero/src/L0Memory.cpp

@@ -174,7 +174,6 @@ MemAllocatorTy::MemPoolTy::MemPoolTy(MemAllocatorTy *_Allocator) {
   BucketStats.resize(1, {0, 0});
   BucketParams.emplace_back(AllocMax, AllocUnit);
   ZeroInit = true;
-  ZeroInitValue.resize(AllocUnit, 0);
   DP("Initialized zero-initialized reduction counter pool for "
      "device " DPxMOD ": AllocMin = %zu, AllocMax = %zu, PoolSizeMax = %zu\n",
      DPxPTR(Allocator->Device), AllocMin, AllocMax, PoolSizeMax);
@@ -266,7 +265,7 @@ void *MemAllocatorTy::MemPoolTy::alloc(size_t Size, size_t &AllocSize) {
 
     if (ZeroInit) {
       auto RC =
-          Allocator->enqueueMemCopy(Base, ZeroInitValue.data(), BlockSize);
+          Allocator->enqueueMemSet(Base, 0, BlockSize);
       if (RC != OFFLOAD_SUCCESS) {
         DP("Failed to zero-initialize pool memory\n");
         return nullptr;
@@ -549,6 +548,10 @@ Error MemAllocatorTy::dealloc_locked(void *Ptr) {
   return Plugin::success();
 }
 
+int32_t MemAllocatorTy::enqueueMemSet(void *Dst, int8_t Value, size_t Size) {
+  return Device->enqueueMemFill(Dst, &Value, sizeof(int8_t), Size);
+}
+
 int32_t MemAllocatorTy::enqueueMemCopy(void *Dst, const void *Src,
                                        size_t Size) {
   return Device->enqueueMemCopy(Dst, Src, Size);

offload/plugins-nextgen/level_zero/src/L0Plugin.cpp

@@ -214,12 +214,8 @@ Error LevelZeroPluginTy::syncBarrierImpl(omp_interop_val_t *Interop) {
        " with ImmCmdList barrier\n",
        DPxPTR(Interop));
     auto ImmCmdList = L0->ImmCmdList;
-    auto Event = l0Device.getEvent();
 
-    CALL_ZE_RET_ERROR(zeCommandListAppendBarrier, ImmCmdList, Event, 0,
-                      nullptr);
-    CALL_ZE_RET_ERROR(zeEventHostSynchronize, Event, UINT64_MAX);
-    l0Device.releaseEvent(Event);
+    CALL_ZE_RET_ERROR(zeCommandListHostSynchronize, ImmCmdList, UINT64_MAX);
   } else {
     DP("LevelZeroPluginTy::sync_barrier: Synchronizing " DPxMOD
        " with queue synchronize\n",
@@ -254,6 +250,8 @@ Error LevelZeroPluginTy::asyncBarrierImpl(omp_interop_val_t *Interop) {
     CALL_ZE_RET_ERROR(zeCommandListAppendBarrier, ImmCmdList, nullptr, 0,
                       nullptr);
   } else {
+#if 0
+    // TODO: re-enable once we have a way to delay the CmdList reset 
     DP("LevelZeroPluginTy::async_barrier: Appending CmdList barrier to " DPxMOD
        "\n",
        DPxPTR(Interop));
@@ -264,6 +262,9 @@ Error LevelZeroPluginTy::asyncBarrierImpl(omp_interop_val_t *Interop) {
     CALL_ZE_RET_ERROR(zeCommandQueueExecuteCommandLists, CmdQueue, 1, &CmdList,
                       nullptr);
     CALL_ZE_RET_ERROR(zeCommandListReset, CmdList);
+#else
+    return syncBarrierImpl(Interop);
+#endif
   }
 
   return Plugin::success();

offload/plugins-nextgen/level_zero/src/L0Program.cpp

@@ -352,7 +352,11 @@ int32_t L0ProgramTy::buildModules(const std::string_view BuildOptions) {
           }
 
           uint64_t Idx = 0;
-          Parts[0].getAsInteger(10, Idx);
+          if (Parts[0].getAsInteger(10, Idx)) {
+            DP("Warning: ignoring auxiliary information (invalid index '%s').\n",
+               Parts[0].str().c_str());
+            continue;
+          }
           MaxImageIdx = (std::max)(MaxImageIdx, Idx);
           if (AuxInfo.find(Idx) != AuxInfo.end()) {
             DP("Warning: duplicate auxiliary information for image %" PRIu64
@@ -362,7 +366,11 @@ int32_t L0ProgramTy::buildModules(const std::string_view BuildOptions) {
           }
 
           uint64_t Part1Id;
-          Parts[1].getAsInteger(10, Part1Id);
+          if (Parts[1].getAsInteger(10, Part1Id)) {
+            DP("Warning: ignoring auxiliary information (invalid part id '%s').\n",
+               Parts[1].str().c_str());
+            continue;
+          }
 
           AuxInfo.emplace(
               std::piecewise_construct, std::forward_as_tuple(Idx),