[L0] don't delete program handle in kernel release and check work dim size

- Decrement the program handle reference count & release the
  L0 resources if the refcnt == 0, but don't delete the program
  handle in the kernel teardown to avoid invalid memory access
  during program handle teardown.

Signed-off-by: Spruit, Neil R <[email protected]>

Author: nrspruit

source/adapters/level_zero/kernel.cpp

@@ -82,10 +82,14 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   uint32_t WG[3]{};
 
   // global_work_size of unused dimensions must be set to 1
-  UR_ASSERT(WorkDim == 3 || GlobalWorkSize[2] == 1,
-            UR_RESULT_ERROR_INVALID_VALUE);
-  UR_ASSERT(WorkDim >= 2 || GlobalWorkSize[1] == 1,
-            UR_RESULT_ERROR_INVALID_VALUE);
+  if (WorkDim >= 2) {
+    UR_ASSERT(WorkDim >= 2 || GlobalWorkSize[1] == 1,
+              UR_RESULT_ERROR_INVALID_VALUE);
+    if (WorkDim == 3) {
+      UR_ASSERT(WorkDim == 3 || GlobalWorkSize[2] == 1,
+                UR_RESULT_ERROR_INVALID_VALUE);
+    }
+  }
   if (LocalWorkSize) {
     // L0
     UR_ASSERT(LocalWorkSize[0] < (std::numeric_limits<uint32_t>::max)(),
@@ -642,8 +646,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelRelease(
   if (IndirectAccessTrackingEnabled) {
     UR_CALL(urContextRelease(KernelProgram->Context));
   }
-  // do a release on the program this kernel was part of
-  UR_CALL(urProgramRelease(KernelProgram));
+  // do a release on the program this kernel was part of without delete of the
+  // program handle
+  KernelProgram->ur_release_program_resources(false);
+
   delete Kernel;
 
   return UR_RESULT_SUCCESS;

source/adapters/level_zero/program.cpp

@@ -837,18 +837,32 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramCreateWithNativeHandle(
 }
 
 ur_program_handle_t_::~ur_program_handle_t_() {
+  if (!resourcesReleased) {
+    ur_release_program_resources(true);
+  }
+}
+
+void ur_program_handle_t_::ur_release_program_resources(bool deletion) {
   // According to Level Zero Specification, all kernels and build logs
   // must be destroyed before the Module can be destroyed.  So, be sure
   // to destroy build log before destroying the module.
-  for (auto &ZeBuildLogPair : this->ZeBuildLogMap) {
-    ZE_CALL_NOCHECK(zeModuleBuildLogDestroy, (ZeBuildLogPair.second));
+  if (!deletion) {
+    if (!RefCount.decrementAndTest()) {
+      return;
+    }
   }
+  if (!resourcesReleased) {
+    for (auto &ZeBuildLogPair : this->ZeBuildLogMap) {
+      ZE_CALL_NOCHECK(zeModuleBuildLogDestroy, (ZeBuildLogPair.second));
+    }
 
-  if (ZeModule && OwnZeModule) {
-    for (auto &ZeModulePair : this->ZeModuleMap) {
-      ZE_CALL_NOCHECK(zeModuleDestroy, (ZeModulePair.second));
+    if (ZeModule && OwnZeModule) {
+      for (auto &ZeModulePair : this->ZeModuleMap) {
+        ZE_CALL_NOCHECK(zeModuleDestroy, (ZeModulePair.second));
+      }
+      this->ZeModuleMap.clear();
     }
-    this->ZeModuleMap.clear();
+    resourcesReleased = true;
   }
 }
 

source/adapters/level_zero/program.hpp

@@ -100,6 +100,11 @@ struct ur_program_handle_t_ : _ur_object {
         State{St}, ZeModule{nullptr}, ZeBuildLog{nullptr} {}
 
   ~ur_program_handle_t_();
+  void ur_release_program_resources(bool deletion);
+
+  // Tracks the release state of the program handle to determine if the
+  // internal handle needs to be released.
+  bool resourcesReleased = false;
 
   const ur_context_handle_t Context; // Context of the program.
 

test/conformance/kernel/kernel_adapter_level_zero.match

@@ -11,7 +11,6 @@ urKernelSetArgMemObjTest.InvalidKernelArgumentIndex/Intel_R__oneAPI_Unified_Runt
 urKernelSetArgPointerTest.SuccessHost/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 urKernelSetArgPointerTest.SuccessDevice/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 urKernelSetArgPointerTest.SuccessShared/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
-urKernelSetArgPointerNegativeTest.InvalidNullHandleKernel/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 urKernelSetArgPointerNegativeTest.InvalidKernelArgumentIndex/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 urKernelSetArgSamplerTest.InvalidKernelArgumentIndex/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 urKernelSetArgValueTest.InvalidKernelArgumentIndex/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_