From a1a224c3bc47617a5b00a2fdeaa4101d3f9cb9bb Mon Sep 17 00:00:00 2001
From: "Hoppe, Mateusz" <mateusz.hoppe@intel.com>
Date: Mon, 21 May 2018 10:57:28 +0200
Subject: [PATCH] Add Clear SLM workaround for BDW

- set ProtectedMemoryDisable in PC prior to setting new L3 config

Change-Id: I2ecbfe9535ef45159b48fffa41faad66db55d074
---
 .../command_stream_receiver_hw.h              |  1 +
 .../command_stream_receiver_hw.inl            |  5 ++++
 runtime/gen8/command_stream_receiver_hw.cpp   |  5 ++++
 .../command_stream_receiver_hw_tests_gen8.cpp | 27 +++++++++++++++++++
 unit_tests/mocks/mock_csr.h                   |  3 +++
 5 files changed, 41 insertions(+)

diff --git a/runtime/command_stream/command_stream_receiver_hw.h b/runtime/command_stream/command_stream_receiver_hw.h
index 8988dc56ec..4a090555e3 100644
--- a/runtime/command_stream/command_stream_receiver_hw.h
+++ b/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();
diff --git a/runtime/command_stream/command_stream_receiver_hw.inl b/runtime/command_stream/command_stream_receiver_hw.inl
index 928e6568c9..ee6814a3b6 100644
--- a/runtime/command_stream/command_stream_receiver_hw.inl
+++ b/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
diff --git a/runtime/gen8/command_stream_receiver_hw.cpp b/runtime/gen8/command_stream_receiver_hw.cpp
index 58c691756a..cdcb502ab4 100644
--- a/runtime/gen8/command_stream_receiver_hw.cpp
+++ b/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>;
 
diff --git a/unit_tests/gen8/command_stream_receiver_hw_tests_gen8.cpp b/unit_tests/gen8/command_stream_receiver_hw_tests_gen8.cpp
index 2c97c3d71d..7689d921c5 100644
--- a/unit_tests/gen8/command_stream_receiver_hw_tests_gen8.cpp
+++ b/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);
+}
diff --git a/unit_tests/mocks/mock_csr.h b/unit_tests/mocks/mock_csr.h
index d7b92635e8..68e8eeb292 100644
--- a/unit_tests/mocks/mock_csr.h
+++ b/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) {}