Add Clear SLM workaround for BDW

- set ProtectedMemoryDisable in PC prior to setting new L3 config

Change-Id: I2ecbfe9535ef45159b48fffa41faad66db55d074

runtime/command_stream/command_stream_receiver_hw.h

@@ -92,6 +92,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
 
     void addPipeControlWA(LinearStream &commandStream, bool flushDC);
     void addDcFlushToPipeControl(typename GfxFamily::PIPE_CONTROL *pCmd, bool flushDC);
+    void addClearSLMWorkAround(typename GfxFamily::PIPE_CONTROL *pCmd);
     PIPE_CONTROL *addPipeControlCmd(LinearStream &commandStream);
 
     uint64_t getScratchPatchAddress();

runtime/command_stream/command_stream_receiver_hw.inl

@@ -653,6 +653,7 @@ inline void CommandStreamReceiverHw<GfxFamily>::programL3(LinearStream &csr, Dis
         *pCmd = GfxFamily::cmdInitPipeControl;
         pCmd->setCommandStreamerStallEnable(true);
         pCmd->setDcFlushEnable(true);
+        addClearSLMWorkAround(pCmd);
 
         PreambleHelper<GfxFamily>::programL3(&csr, newL3Config);
         this->lastSentL3Config = newL3Config;
@@ -727,4 +728,8 @@ void CommandStreamReceiverHw<GfxFamily>::resetKmdNotifyHelper(KmdNotifyHelper *n
     }
 }
 
+template <typename GfxFamily>
+void CommandStreamReceiverHw<GfxFamily>::addClearSLMWorkAround(typename GfxFamily::PIPE_CONTROL *pCmd) {
+}
+
 } // namespace OCLRT

runtime/gen8/command_stream_receiver_hw.cpp

@@ -58,6 +58,11 @@ void populateFactoryTable<CommandStreamReceiverHw<Family>>() {
     commandStreamReceiverFactory[gfxCore] = DeviceCommandStreamReceiver<Family>::create;
 }
 
+template <>
+void CommandStreamReceiverHw<Family>::addClearSLMWorkAround(Family::PIPE_CONTROL *pCmd) {
+    pCmd->setProtectedMemoryDisable(1);
+}
+
 // Explicitly instantiate CommandStreamReceiverHw for this device family
 template class CommandStreamReceiverHw<Family>;
 

unit_tests/gen8/command_stream_receiver_hw_tests_gen8.cpp

@@ -49,3 +49,30 @@ GEN8TEST_F(CommandStreamReceiverHwTestGen8, GivenKernelWithSlmWhenPreviousNOSLML
 GEN8TEST_F(CommandStreamReceiverHwTestGen8, GivenBlockedKernelWithSlmWhenPreviousNOSLML3WasSentOnThenProgramL3WithSLML3ConfigAfterUnblocking) {
     givenBlockedKernelWithSlmWhenPreviousNOSLML3WasSentThenProgramL3WithSLML3ConfigAfterUnblockingImpl();
 }
+
+GEN8TEST_F(CommandStreamReceiverHwTestGen8, GivenChangedL3ConfigWhenL3IsProgrammedThenClearSLMWorkAroundIsAdded) {
+    MockCsrHw2<FamilyType> csr(pDevice->getHardwareInfo());
+    csr.csrSizeRequestFlags.l3ConfigChanged = true;
+    csr.isPreambleSent = true;
+
+    size_t bufferSize = 2 * sizeof(typename FamilyType::MI_LOAD_REGISTER_IMM) + sizeof(typename FamilyType::PIPE_CONTROL);
+    void *buffer = alignedMalloc(bufferSize, 64);
+
+    LinearStream stream(buffer, bufferSize);
+    DispatchFlags flags;
+    uint32_t l3Config = 0x12345678;
+
+    csr.programL3(stream, flags, l3Config);
+
+    this->parseCommands<FamilyType>(stream);
+
+    typename FamilyType::PIPE_CONTROL *pc = getCommand<typename FamilyType::PIPE_CONTROL>();
+    ASSERT_NE(nullptr, pc);
+    EXPECT_TRUE(pc->getProtectedMemoryDisable() != 0);
+
+    typename FamilyType::MI_LOAD_REGISTER_IMM *lri = getCommand<typename FamilyType::MI_LOAD_REGISTER_IMM>();
+    ASSERT_NE(nullptr, lri);
+    EXPECT_EQ(l3Config, lri->getDataDword());
+
+    alignedFree(buffer);
+}

unit_tests/mocks/mock_csr.h

@@ -149,12 +149,15 @@ template <typename GfxFamily>
 class MockCsrHw2 : public CommandStreamReceiverHw<GfxFamily> {
   public:
     using CommandStreamReceiverHw<GfxFamily>::flushStamp;
+    using CommandStreamReceiverHw<GfxFamily>::programL3;
+    using CommandStreamReceiverHw<GfxFamily>::csrSizeRequestFlags;
     using CommandStreamReceiver::commandStream;
     using CommandStreamReceiver::dispatchMode;
     using CommandStreamReceiver::lastSentCoherencyRequest;
     using CommandStreamReceiver::mediaVfeStateDirty;
     using CommandStreamReceiver::taskCount;
     using CommandStreamReceiver::taskLevel;
+    using CommandStreamReceiver::isPreambleSent;
 
     MockCsrHw2(const HardwareInfo &hwInfoIn) : CommandStreamReceiverHw<GfxFamily>(hwInfoIn) {}