Debug flag to force stateless mocs encryption bit

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

Author: BartoszDunajski

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_1_tests.cpp

@@ -122,7 +122,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenCsrInBatch
     mockCsr.getCS(1024u);
     auto &csrCommandStream = mockCsr.commandStream;
 
-    //we do level change that will emit PPC, fill all the space so only BB end fits.
+    // we do level change that will emit PPC, fill all the space so only BB end fits.
     taskLevel++;
     auto ppcSize = MemorySynchronizationCommands<FamilyType>::getSizeForSingleBarrier(false);
     auto fillSize = MemoryConstants::cacheLineSize - ppcSize - sizeof(typename FamilyType::MI_BATCH_BUFFER_END);
@@ -516,7 +516,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, whenSamplerCacheFlushAfterThenSend
 
 HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenStaleCqWhenFlushingTaskThenCompletionStampIsValid) {
     auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
-    //simulate our CQ is stale for 10 TL's
+    // simulate our CQ is stale for 10 TL's
     commandStreamReceiver.taskLevel = taskLevel + 10;
 
     auto completionStamp = flushTask(commandStreamReceiver);
@@ -574,6 +574,25 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, WhenFlushingTas
     EXPECT_EQ(stateHeapMocs, cmd.getInstructionMemoryObjectControlState());
 }
 
+HWTEST_F(CommandStreamReceiverFlushTaskTests, givenDebugVariableSetWhenProgrammingSbaThenSetStatelessMocsEncryptionBit) {
+    using STATE_BASE_ADDRESS = typename FamilyType::STATE_BASE_ADDRESS;
+
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.ForceStatelessMocsEncryptionBit.set(1);
+
+    auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    flushTask(commandStreamReceiver);
+
+    auto &commandStreamCSR = commandStreamReceiver.commandStream;
+    parseCommands<FamilyType>(commandStreamCSR, 0);
+    HardwareParse::findHardwareCommands<FamilyType>();
+
+    ASSERT_NE(nullptr, cmdStateBaseAddress);
+    auto cmd = reinterpret_cast<STATE_BASE_ADDRESS *>(cmdStateBaseAddress);
+
+    EXPECT_EQ(1u, cmd->getStatelessDataPortAccessMemoryObjectControlState() & 1);
+}
+
 HWTEST_F(CommandStreamReceiverFlushTaskTests, givenStateBaseAddressWhenItIsRequiredThenThereIsPipeControlPriorToItWithTextureCacheFlush) {
     typedef typename FamilyType::STATE_BASE_ADDRESS STATE_BASE_ADDRESS;
     auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
@@ -1072,7 +1091,7 @@ HWTEST_F(CommandStreamReceiverCQFlushTaskTests, WhenGettingCsThenReturnCsWithEno
     auto currentUsed = commandStream.getUsed();
     EXPECT_EQ(0u, currentUsed % MemoryConstants::cacheLineSize);
 
-    //depending on the size of commands we may need whole additional cacheline for alignment
+    // depending on the size of commands we may need whole additional cacheline for alignment
     if (currentUsed != expectedSize) {
         EXPECT_EQ(expectedSize - MemoryConstants::cacheLineSize, currentUsed);
     } else {

shared/source/debug_settings/debug_variables_base.inl

@@ -219,6 +219,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideCmdListCmdBufferSizeInKb, -1, "-1: defau
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideL1CachePolicyInSurfaceStateAndStateless, -1, "-1: default, >=0 : following policy will be programmed in render surface state (for regular buffers) and stateless L1 caching")
 DECLARE_DEBUG_VARIABLE(int32_t, PlaformSupportEvictIfNecessaryFlag, -1, "-1: default - platform specific, 0: disable, 1: enable")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceEvictOnlyIfNecessaryFlag, -1, "-1: default - driver selects when to use, 0: force never use this flag, 1: force always use this flag")
+DECLARE_DEBUG_VARIABLE(int32_t, ForceStatelessMocsEncryptionBit, -1, "-1: default - 1: set encryption bit")
 
 /*LOGGING FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, PrintDriverDiagnostics, -1, "prints driver diagnostics messages to standard output, value corresponds to hint level")

shared/source/gen8/state_base_address_gen8.cpp

@@ -64,6 +64,8 @@ void StateBaseAddressHelper<Gen8Family>::programStateBaseAddress(
 
     args.statelessMocsIndex = args.statelessMocsIndex << 1;
 
+    GmmHelper::applyMocsEncryptionBit(args.statelessMocsIndex);
+
     args.stateBaseAddressCmd->setStatelessDataPortAccessMemoryObjectControlState(args.statelessMocsIndex);
 
     appendStateBaseAddressParameters(args, true);

shared/source/gmm_helper/gmm_helper.cpp

@@ -38,6 +38,12 @@ uint32_t GmmHelper::getMOCS(uint32_t type) const {
     return static_cast<uint32_t>(mocs.DwordValue);
 }
 
+void GmmHelper::applyMocsEncryptionBit(uint32_t &index) {
+    if (DebugManager.flags.ForceStatelessMocsEncryptionBit.get() == 1) {
+        index |= 1;
+    }
+}
+
 GmmHelper::GmmHelper(OSInterface *osInterface, const HardwareInfo *pHwInfo) : hwInfo(pHwInfo) {
     auto hwInfoAddressWidth = Math::log2(hwInfo->capabilityTable.gpuAddressSpace + 1);
     addressWidth = std::max(hwInfoAddressWidth, 48u);

shared/source/gmm_helper/gmm_helper.h

@@ -22,6 +22,7 @@ class GmmHelper {
 
     const HardwareInfo *getHardwareInfo();
     uint32_t getMOCS(uint32_t type) const;
+    static void applyMocsEncryptionBit(uint32_t &index);
     void forceAllResourcesUncached() { allResourcesUncached = true; };
 
     static constexpr uint64_t maxPossiblePitch = (1ull << 31);

shared/source/helpers/state_base_address_base.inl

@@ -93,6 +93,8 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
 
     args.statelessMocsIndex = args.statelessMocsIndex << 1;
 
+    GmmHelper::applyMocsEncryptionBit(args.statelessMocsIndex);
+
     args.stateBaseAddressCmd->setStatelessDataPortAccessMemoryObjectControlState(args.statelessMocsIndex);
 
     appendStateBaseAddressParameters(args, overrideBindlessSurfaceStateBase);

shared/source/helpers/state_base_address_xehp_and_later.inl

@@ -86,8 +86,14 @@ void StateBaseAddressHelper<GfxFamily>::appendStateBaseAddressParameters(
         setSbaStatelessCompressionParams<GfxFamily>(args.stateBaseAddressCmd, args.memoryCompressionState);
     }
 
-    if (args.stateBaseAddressCmd->getStatelessDataPortAccessMemoryObjectControlState() == args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) && DebugManager.flags.ForceL1Caching.get() != 0) {
-        args.stateBaseAddressCmd->setStatelessDataPortAccessMemoryObjectControlState(args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CONST));
+    bool l3MocsEnabled = (args.stateBaseAddressCmd->getStatelessDataPortAccessMemoryObjectControlState() >> 1) == (args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) >> 1);
+    bool constMocsAllowed = (l3MocsEnabled && (DebugManager.flags.ForceL1Caching.get() != 0));
+
+    if (constMocsAllowed) {
+        auto constMocsIndex = args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CONST);
+        GmmHelper::applyMocsEncryptionBit(constMocsIndex);
+
+        args.stateBaseAddressCmd->setStatelessDataPortAccessMemoryObjectControlState(constMocsIndex);
     }
 
     appendExtraCacheSettings(args);

shared/test/common/test_files/igdrcl.config

@@ -457,4 +457,5 @@ ForceCsrLockInBcsEnqueueOnlyForGpgpuSubmission = -1
 ExperimentalEnableTileAttach = 0
 DirectSubmissionDisablePrefetcher = -1
 ForceDefaultGrfCompilationMode = 0
-ForceLargeGrfCompilationMode = 0
+ForceLargeGrfCompilationMode = 0
+ForceStatelessMocsEncryptionBit = -1