refactor: Remove template param from enqueueNonBlocked

Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>

Source: bac0207

Author: MaciejPlewka

opencl/source/command_queue/command_queue_hw.h

@@ -367,7 +367,6 @@ class CommandQueueHw : public CommandQueue {
     template <uint32_t cmdType>
     cl_int enqueueBlitSplit(MultiDispatchInfo &dispatchInfo, cl_uint numEventsInWaitList, const cl_event *eventWaitList, cl_event *event, bool blocking, CommandStreamReceiver &csr);
 
-    template <uint32_t commandType>
     CompletionStamp enqueueNonBlocked(Surface **surfacesForResidency,
                                       size_t surfaceCount,
                                       LinearStream &commandStream,
@@ -381,7 +380,8 @@ class CommandQueueHw : public CommandQueue {
                                       EventBuilder &eventBuilder,
                                       TaskCountType taskLevel,
                                       PrintfHandler *printfHandler,
-                                      bool relaxedOrderingEnabled);
+                                      bool relaxedOrderingEnabled,
+                                      uint32_t commandType);
 
     void enqueueBlocked(uint32_t commandType,
                         Surface **surfacesForResidency,

opencl/source/command_queue/enqueue_common.h

@@ -343,7 +343,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
         if (enqueueProperties.operation == EnqueueProperties::Operation::GpuKernel) {
             csrDeps.makeResident(computeCommandStreamReceiver);
 
-            completionStamp = enqueueNonBlocked<commandType>(
+            completionStamp = enqueueNonBlocked(
                 surfacesForResidency,
                 numSurfaceForResidency,
                 commandStream,
@@ -357,7 +357,8 @@ cl_int CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
                 eventBuilder,
                 taskLevel,
                 printfHandler.get(),
-                relaxedOrderingEnabled);
+                relaxedOrderingEnabled,
+                commandType);
         } else if (enqueueProperties.isFlushWithoutKernelRequired()) {
             completionStamp = enqueueCommandWithoutKernel(
                 surfacesForResidency,
@@ -745,7 +746,6 @@ bool CommandQueueHw<GfxFamily>::isTaskLevelUpdateRequired(const TaskCountType &t
 }
 
 template <typename GfxFamily>
-template <uint32_t commandType>
 CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     Surface **surfaces,
     size_t surfaceCount,
@@ -760,7 +760,8 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
     EventBuilder &eventBuilder,
     TaskCountType taskLevel,
     PrintfHandler *printfHandler,
-    bool relaxedOrderingEnabled) {
+    bool relaxedOrderingEnabled,
+    uint32_t commandType) {
 
     UNRECOVERABLE_IF(multiDispatchInfo.empty());
 

opencl/test/unit_test/xe_hpg_core/dg2/command_queue_tests_dg2.cpp

@@ -80,8 +80,8 @@ DG2TEST_F(CommandQueueHwTest, GivenKernelWithDpasAndOddWorkGroupWhenenqueueNonBl
 
     bool blocking = false;
     pKernel->setSystolicPipelineSelectMode(true);
-    cmdQ.template enqueueNonBlocked<CL_COMMAND_NDRANGE_KERNEL>(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
-                                                               multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false);
+    cmdQ.enqueueNonBlocked(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
+                           multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false, CL_COMMAND_NDRANGE_KERNEL);
     EXPECT_TRUE(csr->disableEuFusionPassed);
 }
 
@@ -117,8 +117,8 @@ DG2TEST_F(CommandQueueHwTest, GivenKernelWithDpasAndNotOddWorkGroupWhenenqueueNo
 
     bool blocking = false;
     const_cast<NEO::KernelDescriptor &>(pKernel->getDescriptor()).kernelAttributes.flags.usesSystolicPipelineSelectMode = true;
-    cmdQ.template enqueueNonBlocked<CL_COMMAND_NDRANGE_KERNEL>(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
-                                                               multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false);
+    cmdQ.enqueueNonBlocked(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
+                           multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false, CL_COMMAND_NDRANGE_KERNEL);
     EXPECT_FALSE(csr->disableEuFusionPassed);
 }
 DG2TEST_F(CommandQueueHwTest, GivenKernelWithRequiredDisableEuFusionWhenenqueueNonBlockedCalledThenDisableEuFusionPassedToFlushTask) {
@@ -144,8 +144,8 @@ DG2TEST_F(CommandQueueHwTest, GivenKernelWithRequiredDisableEuFusionWhenenqueueN
 
     bool blocking = false;
     const_cast<NEO::KernelDescriptor &>(pKernel->getDescriptor()).kernelAttributes.flags.requiresDisabledEUFusion = true;
-    cmdQ.template enqueueNonBlocked<CL_COMMAND_NDRANGE_KERNEL>(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
-                                                               multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false);
+    cmdQ.enqueueNonBlocked(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true,
+                           multiDispatchInfo, enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false, CL_COMMAND_NDRANGE_KERNEL);
     EXPECT_TRUE(csr->disableEuFusionPassed);
 }
 DG2TEST_F(CommandQueueHwTest, GivenKernelWithoutRequiredDisableEuFusionWhenenqueueNonBlockedCalledThenDisableEuFusionNotPassedToFlushTask) {
@@ -171,7 +171,7 @@ DG2TEST_F(CommandQueueHwTest, GivenKernelWithoutRequiredDisableEuFusionWhenenque
 
     bool blocking = false;
     const_cast<NEO::KernelDescriptor &>(pKernel->getDescriptor()).kernelAttributes.flags.requiresDisabledEUFusion = false;
-    cmdQ.template enqueueNonBlocked<CL_COMMAND_NDRANGE_KERNEL>(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true, multiDispatchInfo,
-                                                               enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false);
+    cmdQ.enqueueNonBlocked(nullptr, 0, commandStream, commandStream.getUsed(), blocking, true, multiDispatchInfo,
+                           enqueueProperties, timestampPacketDependencies, eventsRequest, eventBuilder, 0, nullptr, false, CL_COMMAND_NDRANGE_KERNEL);
     EXPECT_FALSE(csr->disableEuFusionPassed);
 }