[SYCL][L0] Cache the result of zeKernelGetName (#6053)

Signed-off-by: Sergey V Maslov <[email protected]>

Author: smaslov-intel

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -4636,18 +4636,31 @@ pi_result piKernelCreate(pi_program Program, const char *KernelName,
     return PI_ERROR_UNKNOWN;
   }
 
-  // Update the refcount of the program and context to show it's used by this
-  // kernel.
+  PI_CALL((*RetKernel)->initialize());
+  return PI_SUCCESS;
+}
+
+pi_result _pi_kernel::initialize() {
+  // Retain the program and context to show it's used by this kernel.
   PI_CALL(piProgramRetain(Program));
   if (IndirectAccessTrackingEnabled)
     // TODO: do piContextRetain without the guard
     PI_CALL(piContextRetain(Program->Context));
 
   // Set up how to obtain kernel properties when needed.
-  (*RetKernel)->ZeKernelProperties.Compute =
-      [ZeKernel](ze_kernel_properties_t &Properties) {
-        ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
-      };
+  ZeKernelProperties.Compute = [this](ze_kernel_properties_t &Properties) {
+    ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
+  };
+
+  // Cache kernel name.
+  ZeKernelName.Compute = [this](std::string &Name) {
+    size_t Size = 0;
+    ZE_CALL_NOCHECK(zeKernelGetName, (ZeKernel, &Size, nullptr));
+    char *KernelName = new char[Size];
+    ZE_CALL_NOCHECK(zeKernelGetName, (ZeKernel, &Size, KernelName));
+    Name = KernelName;
+    delete[] KernelName;
+  };
 
   return PI_SUCCESS;
 }
@@ -4727,13 +4740,8 @@ pi_result piKernelGetInfo(pi_kernel Kernel, pi_kernel_info ParamName,
     return ReturnValue(pi_program{Kernel->Program});
   case PI_KERNEL_INFO_FUNCTION_NAME:
     try {
-      size_t Size = 0;
-      ZE_CALL(zeKernelGetName, (Kernel->ZeKernel, &Size, nullptr));
-      char *KernelName = new char[Size];
-      ZE_CALL(zeKernelGetName, (Kernel->ZeKernel, &Size, KernelName));
-      pi_result Res = ReturnValue(static_cast<const char *>(KernelName));
-      delete[] KernelName;
-      return Res;
+      std::string &KernelName = *Kernel->ZeKernelName.operator->();
+      return ReturnValue(static_cast<const char *>(KernelName.c_str()));
     } catch (const std::bad_alloc &) {
       return PI_OUT_OF_HOST_MEMORY;
     } catch (...) {
@@ -5086,20 +5094,7 @@ pi_result piextKernelCreateWithNativeHandle(pi_native_handle NativeHandle,
 
   auto ZeKernel = pi_cast<ze_kernel_handle_t>(NativeHandle);
   *Kernel = new _pi_kernel(ZeKernel, OwnNativeHandle, Program);
-
-  // Update the refcount of the program and context to show it's used by this
-  // kernel.
-  PI_CALL(piProgramRetain(Program));
-  if (IndirectAccessTrackingEnabled)
-    // TODO: do piContextRetain without the guard
-    PI_CALL(piContextRetain(Program->Context));
-
-  // Set up how to obtain kernel properties when needed.
-  (*Kernel)->ZeKernelProperties.Compute =
-      [ZeKernel](ze_kernel_properties_t &Properties) {
-        ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
-      };
-
+  PI_CALL((*Kernel)->initialize());
   return PI_SUCCESS;
 }
 

sycl/plugins/level_zero/pi_level_zero.hpp

@@ -185,7 +185,7 @@ template <class T> struct ZeCache : private T {
   // order to disallow access other than through "->".
   //
   typedef std::function<void(T &)> InitFunctionType;
-  InitFunctionType Compute;
+  InitFunctionType Compute{nullptr};
   bool Computed{false};
 
   ZeCache() : T{} {}
@@ -1379,6 +1379,9 @@ struct _pi_kernel : _pi_object {
       : ZeKernel{Kernel}, OwnZeKernel{OwnZeKernel}, Program{Program},
         MemAllocs{}, SubmissionsCount{0} {}
 
+  // Completed initialization of PI kernel. Must be called after construction.
+  pi_result initialize();
+
   // Returns true if kernel has indirect access, false otherwise.
   bool hasIndirectAccess() {
     // Currently indirect access flag is set for all kernels and there is no API
@@ -1445,6 +1448,7 @@ struct _pi_kernel : _pi_object {
 
   // Cache of the kernel properties.
   ZeCache<ZeStruct<ze_kernel_properties_t>> ZeKernelProperties;
+  ZeCache<std::string> ZeKernelName;
 };
 
 struct _pi_sampler : _pi_object {