Add WaDisableFusedThreadScheduling

Change-Id: I5ad09fc7d366b8062ff7b10f86718f3afe28ba0b
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>

Author: KamilKoprykIntel

core/gen12lp/preamble_gen12lp.cpp

@@ -76,7 +76,8 @@ uint32_t PreambleHelper<TGLLPFamily>::getUrbEntryAllocationSize() {
 }
 
 template <>
-void PreambleHelper<TGLLPFamily>::programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState) {
+void PreambleHelper<TGLLPFamily>::programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState, const HardwareInfo &hwInfo) {
+    mediaVfeState->setDisableSlice0Subslice2(hwInfo.workaroundTable.waDisableFusedThreadScheduling);
 
     if (DebugManager.flags.CFEFusedEUDispatch.get() != -1) {
         mediaVfeState->setDisableSlice0Subslice2(DebugManager.flags.CFEFusedEUDispatch.get());

core/helpers/preamble.h

@@ -41,7 +41,7 @@ struct PreambleHelper {
                                     int scratchSize,
                                     uint64_t scratchAddress,
                                     uint32_t maxFrontEndThreads);
-    static void programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState);
+    static void programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState, const HardwareInfo &hwInfo);
     static void programPreamble(LinearStream *pCommandStream, Device &device, uint32_t l3Config,
                                 uint32_t requiredThreadArbitrationPolicy, GraphicsAllocation *preemptionCsr, GraphicsAllocation *perDssBackedBuffer);
     static void programKernelDebugging(LinearStream *pCommandStream);

core/helpers/preamble_base.inl

@@ -112,7 +112,7 @@ bool PreambleHelper<GfxFamily>::isL3Configurable(const HardwareInfo &hwInfo) {
 }
 
 template <typename GfxFamily>
-void PreambleHelper<GfxFamily>::programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState) {
+void PreambleHelper<GfxFamily>::programAdditionalFieldsInVfeState(VFE_STATE_TYPE *mediaVfeState, const HardwareInfo &hwInfo) {
 }
 
 } // namespace NEO

core/helpers/preamble_bdw_plus.inl

@@ -48,7 +48,7 @@ uint64_t PreambleHelper<GfxFamily>::programVFEState(LinearStream *pCommandStream
     pMediaVfeState->setScratchSpaceBasePointer(lowAddress);
     pMediaVfeState->setScratchSpaceBasePointerHigh(highAddress);
 
-    programAdditionalFieldsInVfeState(pMediaVfeState);
+    programAdditionalFieldsInVfeState(pMediaVfeState, hwInfo);
 
     return scratchSpaceAddressOffset;
 }

runtime/sku_info/operations/sku_info_receiver.h

@@ -129,6 +129,7 @@ class SkuInfoReceiver {
         RECEIVE_WA(AuxTable16KGranular);
         RECEIVE_WA(Limit128BMediaCompr);
         RECEIVE_WA(UntypedBufferCompression);
+        RECEIVE_WA(DisableFusedThreadScheduling);
 #undef RECEIVE_WA
     }
 };

runtime/sku_info/sku_info_base.h

@@ -117,5 +117,6 @@ struct WorkaroundTableBase {
     bool waLimit128BMediaCompr = false;
     bool waUntypedBufferCompression = false;
     bool waAuxTable16KGranular = false;
+    bool waDisableFusedThreadScheduling = false;
 };
 } // namespace NEO

unit_tests/gen12lp/test_preamble_gen12lp.cpp

@@ -129,20 +129,28 @@ TGLLPTEST_F(Gen12LpPreambleVfeState, givenDefaultPipeControlWhenItIsProgrammedTh
 TGLLPTEST_F(Gen12LpPreambleVfeState, givenCfeFusedEuDispatchFlagsWhenprogramAdditionalFieldsInVfeStateIsCalledThenGetDisableSlice0Subslice2ReturnsCorrectValues) {
     using MEDIA_VFE_STATE = typename FamilyType::MEDIA_VFE_STATE;
 
+    DebugManagerStateRestore restorer;
+    auto pHwInfo = pPlatform->getDevice(0)->getExecutionEnvironment()->getMutableHardwareInfo();
     auto pMediaVfeState = reinterpret_cast<MEDIA_VFE_STATE *>(linearStream.getSpace(sizeof(MEDIA_VFE_STATE)));
     *pMediaVfeState = FamilyType::cmdInitMediaVfeState;
-
-    DebugManagerStateRestore restorer;
-
-    std::vector<std::pair<bool, int32_t>> testParams{{false, 0},
-                                                     {false, -1},
-                                                     {true, 1},
-                                                     {true, -1}};
-
-    for (const auto &it : testParams) {
-        ::DebugManager.flags.CFEFusedEUDispatch.set(it.second);
-        PreambleHelper<FamilyType>::programAdditionalFieldsInVfeState(pMediaVfeState);
-        EXPECT_EQ(it.first, pMediaVfeState->getDisableSlice0Subslice2());
+    auto &waTable = pHwInfo->workaroundTable;
+
+    const std::array<std::tuple<bool, bool, int32_t>, 6> testParams{std::make_tuple(false, false, 0),
+                                                                    std::make_tuple(false, true, 0),
+                                                                    std::make_tuple(false, false, -1),
+                                                                    std::make_tuple(true, false, 1),
+                                                                    std::make_tuple(true, true, -1),
+                                                                    std::make_tuple(true, true, 1)};
+
+    for (const auto &params : testParams) {
+        bool expectedValue, waDisableFusedThreadScheduling;
+        int32_t debugKeyValue;
+        std::tie(expectedValue, waDisableFusedThreadScheduling, debugKeyValue) = params;
+
+        waTable.waDisableFusedThreadScheduling = waDisableFusedThreadScheduling;
+        ::DebugManager.flags.CFEFusedEUDispatch.set(debugKeyValue);
+        PreambleHelper<FamilyType>::programAdditionalFieldsInVfeState(pMediaVfeState, *pHwInfo);
+        EXPECT_EQ(expectedValue, pMediaVfeState->getDisableSlice0Subslice2());
     }
 }
 

unit_tests/sku_info/sku_info_base_reference.h

@@ -165,6 +165,7 @@ struct SkuInfoBaseReference {
         refWaTable.waLimit128BMediaCompr = 1;
         refWaTable.waUntypedBufferCompression = 1;
         refWaTable.waAuxTable16KGranular = 1;
+        refWaTable.waDisableFusedThreadScheduling = true;
     }
 }; // namespace SkuInfoBaseReference
 } // namespace NEO