Limit number of pipeline select commands when using multiple command queues

Related-To: NEO-7187

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

level_zero/core/source/cmdqueue/cmdqueue_hw.inl

@@ -251,6 +251,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
 
     gsbaStateDirty |= csr->getGSBAStateDirty();
     frontEndStateDirty |= csr->getMediaVFEStateDirty();
+    bool gpgpuEnabled = csr->getPreambleSetFlag();
     if (!isCopyOnlyCommandQueue) {
 
         if (!gpgpuEnabled) {
@@ -315,6 +316,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
     if (!isCopyOnlyCommandQueue) {
         if (!gpgpuEnabled) {
             programPipelineSelect(child);
+            csr->setPreambleSetFlag(true);
         }
 
         if (NEO::Debugger::isDebugEnabled(internalUsage) && !commandQueueDebugCmdsProgrammed) {
@@ -585,7 +587,6 @@ template <GFXCORE_FAMILY gfxCoreFamily>
 void CommandQueueHw<gfxCoreFamily>::programPipelineSelect(NEO::LinearStream &commandStream) {
     NEO::PipelineSelectArgs args = {0, 0};
     NEO::PreambleHelper<GfxFamily>::programPipelineSelect(&commandStream, args, device->getHwInfo());
-    gpgpuEnabled = true;
 }
 
 template <GFXCORE_FAMILY gfxCoreFamily>

level_zero/core/source/cmdqueue/cmdqueue_imp.h

@@ -101,7 +101,6 @@ struct CommandQueueImp : public CommandQueue {
 
     std::atomic<uint32_t> taskCount{0};
 
-    bool gpgpuEnabled = false;
     bool useKmdWaitFunction = false;
 };
 

level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue_enqueue_cmdlist_2.cpp

@@ -121,5 +121,67 @@ HWTEST2_F(CommandQueueExecuteCommandListsSimpleTest, whenUsingFenceThenLastPipeC
     commandQueue->destroy();
 }
 
+HWTEST2_F(CommandQueueExecuteCommandListsSimpleTest, givenTwoCommandQueuesUsingSingleCsrWhenExecutingFirstTimeOnBothThenPipelineSelectProgrammedOnce, IsAtMostXeHpcCore) {
+    using PIPELINE_SELECT = typename FamilyType::PIPELINE_SELECT;
+
+    bool additionalPipelineSelect = NEO::HwInfoConfig::get(device->getHwInfo().platform.eProductFamily)->is3DPipelineSelectWARequired() &&
+                                    neoDevice->getDefaultEngine().commandStreamReceiver->isRcs();
+
+    ze_result_t returnValue;
+
+    ze_command_list_handle_t commandList = CommandList::create(productFamily, device, NEO::EngineGroupType::RenderCompute, 0u, returnValue)->toHandle();
+    ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
+
+    ze_command_queue_desc_t queueDesc = {};
+    queueDesc.mode = ZE_COMMAND_QUEUE_MODE_ASYNCHRONOUS;
+    auto commandQueue = whiteboxCast(CommandQueue::create(productFamily, device, neoDevice->getDefaultEngine().commandStreamReceiver, &queueDesc, false, false, returnValue));
+    ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
+    ASSERT_NE(nullptr, commandQueue->commandStream);
+
+    auto usedSpaceBefore = commandQueue->commandStream->getUsed();
+    returnValue = commandQueue->executeCommandLists(1, &commandList, nullptr, false);
+    ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
+    auto usedSpaceAfter = commandQueue->commandStream->getUsed();
+    ASSERT_GT(usedSpaceAfter, usedSpaceBefore);
+
+    GenCmdList cmdList;
+    ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
+        cmdList,
+        ptrOffset(commandQueue->commandStream->getCpuBase(), usedSpaceBefore),
+        usedSpaceAfter - usedSpaceBefore));
+
+    auto pipelineSelect = findAll<PIPELINE_SELECT *>(cmdList.begin(), cmdList.end());
+    size_t expectedFirstPipelineSelectCount = 1u;
+    if (additionalPipelineSelect) {
+        expectedFirstPipelineSelectCount += 2;
+    }
+    EXPECT_EQ(expectedFirstPipelineSelectCount, pipelineSelect.size());
+
+    cmdList.clear();
+
+    auto commandQueue2 = whiteboxCast(CommandQueue::create(productFamily, device, neoDevice->getDefaultEngine().commandStreamReceiver, &queueDesc, false, false, returnValue));
+    ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
+    ASSERT_NE(nullptr, commandQueue2->commandStream);
+
+    usedSpaceBefore = commandQueue2->commandStream->getUsed();
+    returnValue = commandQueue2->executeCommandLists(1, &commandList, nullptr, false);
+    ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
+    usedSpaceAfter = commandQueue2->commandStream->getUsed();
+    ASSERT_GT(usedSpaceAfter, usedSpaceBefore);
+
+    ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
+        cmdList,
+        ptrOffset(commandQueue2->commandStream->getCpuBase(), usedSpaceBefore),
+        usedSpaceAfter - usedSpaceBefore));
+
+    pipelineSelect = findAll<PIPELINE_SELECT *>(cmdList.begin(), cmdList.end());
+    constexpr size_t expectedSecondPipelineSelectCount = 0u;
+    EXPECT_EQ(expectedSecondPipelineSelectCount, pipelineSelect.size());
+
+    CommandList::fromHandle(commandList)->destroy();
+    commandQueue->destroy();
+    commandQueue2->destroy();
+}
+
 } // namespace ult
 } // namespace L0

shared/source/command_stream/command_stream_receiver.h

@@ -340,6 +340,14 @@ class CommandStreamReceiver {
 
     LogicalStateHelper *getLogicalStateHelper() const;
 
+    bool getPreambleSetFlag() const {
+        return isPreambleSent;
+    }
+
+    void setPreambleSetFlag(bool value) {
+        isPreambleSent = value;
+    }
+
   protected:
     void cleanupResources();
     void printDeviceIndex();

shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp

@@ -2097,3 +2097,9 @@ HWTEST_F(CommandStreamReceiverTest, givenMultipleActivePartitionsWhenWaitLogIsEn
 
     EXPECT_STREQ(expectedOutput.str().c_str(), output.c_str());
 }
+
+TEST_F(CommandStreamReceiverTest, givenPreambleFlagIsSetWhenGettingFlagStateThenExpectCorrectState) {
+    EXPECT_FALSE(commandStreamReceiver->getPreambleSetFlag());
+    commandStreamReceiver->setPreambleSetFlag(true);
+    EXPECT_TRUE(commandStreamReceiver->getPreambleSetFlag());
+}