fix: wa for multi ccs execution on bmg

Related-To: NEO-15771
Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>

opencl/test/unit_test/xe2_hpg_core/test_preamble_xe2_hpg_core.cpp

@@ -27,7 +27,7 @@ XE2_HPG_CORETEST_F(PreambleCfeState, givenXe2HpgCoreAndConcurrentKernelExecution
     EXPECT_FALSE(cfeState->getSingleSliceDispatchCcsMode());
 }
 
-XE2_HPG_CORETEST_F(PreambleCfeState, givenXe2HpgCoreAndDefaultKernelExecutionTypeWhenCallingProgramVFEStateThenSingleSpliceDispatchCcsModeIsDisabled) {
+XE2_HPG_CORETEST_F(PreambleCfeState, givenXe2HpgCoreAndDefaultKernelExecutionTypeWhenCallingProgramVFEStateThenSingleSliceDispatchCcsModeIsDisabled) {
     using CFE_STATE = typename FamilyType::CFE_STATE;
 
     auto pVfeCmd = PreambleHelper<FamilyType>::getSpaceForVfeState(&linearStream, *defaultHwInfo, EngineGroupType::renderCompute, nullptr);
@@ -42,3 +42,21 @@ XE2_HPG_CORETEST_F(PreambleCfeState, givenXe2HpgCoreAndDefaultKernelExecutionTyp
     EXPECT_FALSE(cfeState->getComputeOverdispatchDisable());
     EXPECT_FALSE(cfeState->getSingleSliceDispatchCcsMode());
 }
+
+BMGTEST_F(PreambleCfeState, givenXe2HpgCoreWithMultiCCSAndDefaultKernelExecutionTypeWhenCallingProgramVFEStateThenSingleSliceDispatchCcsModeIsEnabled) {
+    using CFE_STATE = typename FamilyType::CFE_STATE;
+
+    auto mutableHwInfo = pDevice->getRootDeviceEnvironment().getMutableHardwareInfo();
+    mutableHwInfo->gtSystemInfo.CCSInfo.NumberOfCCSEnabled = 2;
+    auto pVfeCmd = PreambleHelper<FamilyType>::getSpaceForVfeState(&linearStream, *defaultHwInfo, EngineGroupType::renderCompute, nullptr);
+    StreamProperties streamProperties{};
+    streamProperties.initSupport(pDevice->getRootDeviceEnvironment());
+    streamProperties.frontEndState.setPropertiesAll(false, false, false);
+    PreambleHelper<FamilyType>::programVfeState(pVfeCmd, pDevice->getRootDeviceEnvironment(), 0u, 0, 0, streamProperties);
+    parseCommands<FamilyType>(linearStream);
+    auto cfeStateIt = find<CFE_STATE *>(cmdList.begin(), cmdList.end());
+    ASSERT_NE(cmdList.end(), cfeStateIt);
+    auto cfeState = reinterpret_cast<CFE_STATE *>(*cfeStateIt);
+    EXPECT_FALSE(cfeState->getComputeOverdispatchDisable());
+    EXPECT_TRUE(cfeState->getSingleSliceDispatchCcsMode());
+}

shared/source/execution_environment/root_device_environment.cpp

@@ -247,7 +247,6 @@ BuiltIns *RootDeviceEnvironment::getBuiltIns() {
 }
 
 void RootDeviceEnvironment::setNumberOfCcs(uint32_t numberOfCcs) {
-
     hwInfo->gtSystemInfo.CCSInfo.NumberOfCCSEnabled = std::min(hwInfo->gtSystemInfo.CCSInfo.NumberOfCCSEnabled, numberOfCcs);
     limitedNumberOfCcs = true;
     if (aubCenter) {
@@ -255,6 +254,10 @@ void RootDeviceEnvironment::setNumberOfCcs(uint32_t numberOfCcs) {
     }
 }
 
+uint32_t RootDeviceEnvironment::getNumberOfCcs() const {
+    return hwInfo->gtSystemInfo.CCSInfo.NumberOfCCSEnabled;
+}
+
 bool RootDeviceEnvironment::isNumberOfCcsLimited() const {
     return limitedNumberOfCcs;
 }

shared/source/execution_environment/root_device_environment.h

@@ -82,6 +82,7 @@ struct RootDeviceEnvironment : NonCopyableClass {
     AssertHandler *getAssertHandler(Device *neoDevice);
     void createBindlessHeapsHelper(Device *rootDevice, bool availableDevices);
     void setNumberOfCcs(uint32_t numberOfCcs);
+    uint32_t getNumberOfCcs() const;
     bool isNumberOfCcsLimited() const;
     void setRcsExposure();
     void initProductHelper();

shared/source/helpers/preamble_xehp_and_later.inl

@@ -13,6 +13,7 @@
 #include "shared/source/helpers/pipe_control_args.h"
 #include "shared/source/helpers/pipeline_select_helper.h"
 #include "shared/source/helpers/preamble_base.inl"
+#include "shared/source/release_helper/release_helper.h"
 
 namespace NEO {
 
@@ -64,8 +65,9 @@ void PreambleHelper<GfxFamily>::programVfeState(void *pVfeState,
         cmd.setMaximumNumberOfThreads(maxFrontEndThreads);
 
         cmd.setComputeOverdispatchDisable(streamProperties.frontEndState.disableOverdispatch.value == 1);
+        auto singleSliceDispatchCcsMode = streamProperties.frontEndState.singleSliceDispatchCcsMode.value == 1 || (rootDeviceEnvironment.getNumberOfCcs() > 1 && rootDeviceEnvironment.getReleaseHelper()->isSingleDispatchRequiredForMultiCCS());
 
-        PreambleHelper<GfxFamily>::setSingleSliceDispatchMode(&cmd, streamProperties.frontEndState.singleSliceDispatchCcsMode.value == 1);
+        PreambleHelper<GfxFamily>::setSingleSliceDispatchMode(&cmd, singleSliceDispatchCcsMode);
 
         appendProgramVFEState(rootDeviceEnvironment, streamProperties, &cmd);
 

shared/source/release_helper/release_helper.h

@@ -70,6 +70,7 @@ class ReleaseHelper {
     virtual uint32_t adjustMaxThreadsPerEuCount(uint32_t maxThreadsPerEuCount, uint32_t grfCount) const = 0;
     virtual bool shouldQueryPeerAccess() const = 0;
     virtual bool isSpirSupported() const = 0;
+    virtual bool isSingleDispatchRequiredForMultiCCS() const = 0;
 
   protected:
     ReleaseHelper(HardwareIpVersion hardwareIpVersion) : hardwareIpVersion(hardwareIpVersion) {}
@@ -119,6 +120,7 @@ class ReleaseHelperHw : public ReleaseHelper {
     uint32_t adjustMaxThreadsPerEuCount(uint32_t maxThreadsPerEuCount, uint32_t grfCount) const override;
     bool shouldQueryPeerAccess() const override;
     bool isSpirSupported() const override;
+    bool isSingleDispatchRequiredForMultiCCS() const override;
 
   protected:
     ReleaseHelperHw(HardwareIpVersion hardwareIpVersion) : ReleaseHelper(hardwareIpVersion) {}

shared/source/release_helper/release_helper_2001.cpp

@@ -54,6 +54,11 @@ bool ReleaseHelperHw<release>::shouldQueryPeerAccess() const {
     return true;
 }
 
+template <>
+bool ReleaseHelperHw<release>::isSingleDispatchRequiredForMultiCCS() const {
+    return true;
+}
+
 } // namespace NEO
 
 #include "shared/source/release_helper/release_helper_common_xe2_hpg.inl"

shared/source/release_helper/release_helper_2002.cpp

@@ -54,6 +54,11 @@ bool ReleaseHelperHw<release>::shouldQueryPeerAccess() const {
     return true;
 }
 
+template <>
+bool ReleaseHelperHw<release>::isSingleDispatchRequiredForMultiCCS() const {
+    return true;
+}
+
 } // namespace NEO
 
 #include "shared/source/release_helper/release_helper_common_xe2_hpg.inl"

shared/source/release_helper/release_helper_base.inl

@@ -196,4 +196,9 @@ bool ReleaseHelperHw<releaseType>::isSpirSupported() const {
     return true;
 }
 
+template <ReleaseType releaseType>
+bool ReleaseHelperHw<releaseType>::isSingleDispatchRequiredForMultiCCS() const {
+    return false;
+}
+
 } // namespace NEO

shared/source/release_helper/release_helper_common_xe2_hpg.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2024 Intel Corporation
+ * Copyright (C) 2024-2025 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,5 +32,4 @@ template <>
 bool ReleaseHelperHw<release>::isGlobalBindlessAllocatorEnabled() const {
     return true;
 }
-
 } // namespace NEO

shared/test/common/mocks/mock_release_helper.h

@@ -48,6 +48,7 @@ class MockReleaseHelper : public ReleaseHelper {
     ADDMETHOD_CONST_NOBASE_VOIDRETURN(adjustRTDispatchGlobals, (void *rtDispatchGlobals, uint32_t rtStacksPerDss, bool heaplessEnabled, uint32_t maxBvhLevels));
     ADDMETHOD_CONST_NOBASE(shouldQueryPeerAccess, bool, false, ());
     ADDMETHOD_CONST_NOBASE(isSpirSupported, bool, true, ());
+    ADDMETHOD_CONST_NOBASE(isSingleDispatchRequiredForMultiCCS, bool, false, ());
 
     const SizeToPreferredSlmValueArray &getSizeToPreferredSlmValue(bool isHeapless) const override {
         static SizeToPreferredSlmValueArray sizeToPreferredSlmValue = {};

shared/test/unit_test/release_helper/release_helper_12_55_tests.cpp

@@ -126,3 +126,7 @@ TEST_F(ReleaseHelper1255Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1255Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1255Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_56_tests.cpp

@@ -126,3 +126,7 @@ TEST_F(ReleaseHelper1256Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1256Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1256Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_57_tests.cpp

@@ -126,3 +126,7 @@ TEST_F(ReleaseHelper1257Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1257Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1257Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_60_tests.cpp

@@ -110,3 +110,7 @@ TEST_F(ReleaseHelper1260Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1260Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1260Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_61_tests.cpp

@@ -109,3 +109,7 @@ TEST_F(ReleaseHelper1261Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1261Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1261Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_70_tests.cpp

@@ -120,3 +120,7 @@ TEST_F(ReleaseHelper1270Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1270Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1270Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_71_tests.cpp

@@ -120,3 +120,7 @@ TEST_F(ReleaseHelper1271Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1271Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1271Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_12_74_tests.cpp

@@ -126,3 +126,7 @@ TEST_F(ReleaseHelper1274Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper1274Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper1274Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_20_01_tests.cpp

@@ -119,3 +119,7 @@ TEST_F(ReleaseHelper2001Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper2001Tests, whenShouldQueryPeerAccessCalledThenTrueReturned) {
     whenShouldQueryPeerAccessCalledThenTrueReturned();
 }
+
+TEST_F(ReleaseHelper2001Tests, whenIsSingleDispatchRequiredForMultiCCSThenTrueReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenTrueReturned();
+}

shared/test/unit_test/release_helper/release_helper_20_02_tests.cpp

@@ -119,3 +119,7 @@ TEST_F(ReleaseHelper2002Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper2002Tests, whenShouldQueryPeerAccessCalledThenTrueReturned) {
     whenShouldQueryPeerAccessCalledThenTrueReturned();
 }
+
+TEST_F(ReleaseHelper2002Tests, whenIsSingleDispatchRequiredForMultiCCSThenTrueReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenTrueReturned();
+}

shared/test/unit_test/release_helper/release_helper_20_04_tests.cpp

@@ -122,3 +122,7 @@ TEST_F(ReleaseHelper2004Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper2004Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper2004Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_30_00_tests.cpp

@@ -120,3 +120,7 @@ TEST_F(ReleaseHelper3000Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper3000Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper3000Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_30_01_tests.cpp

@@ -120,3 +120,7 @@ TEST_F(ReleaseHelper3001Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper3001Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper3001Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_30_03_tests.cpp

@@ -120,3 +120,7 @@ TEST_F(ReleaseHelper3003Tests, whenCallingAdjustMaxThreadsPerEuCountThenCorrectV
 TEST_F(ReleaseHelper3003Tests, whenShouldQueryPeerAccessCalledThenFalseReturned) {
     whenShouldQueryPeerAccessCalledThenFalseReturned();
 }
+
+TEST_F(ReleaseHelper3003Tests, whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned) {
+    whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+}

shared/test/unit_test/release_helper/release_helper_tests_base.cpp

@@ -214,3 +214,21 @@ void ReleaseHelperTestsBase::whenShouldQueryPeerAccessCalledThenTrueReturned() {
         EXPECT_TRUE(releaseHelper->shouldQueryPeerAccess());
     }
 }
+
+void ReleaseHelperTestsBase::whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned() {
+    for (auto &revision : getRevisions()) {
+        ipVersion.revision = revision;
+        releaseHelper = ReleaseHelper::create(ipVersion);
+        ASSERT_NE(nullptr, releaseHelper);
+        EXPECT_FALSE(releaseHelper->isSingleDispatchRequiredForMultiCCS());
+    }
+}
+
+void ReleaseHelperTestsBase::whenIsSingleDispatchRequiredForMultiCCSCalledThenTrueReturned() {
+    for (auto &revision : getRevisions()) {
+        ipVersion.revision = revision;
+        releaseHelper = ReleaseHelper::create(ipVersion);
+        ASSERT_NE(nullptr, releaseHelper);
+        EXPECT_TRUE(releaseHelper->isSingleDispatchRequiredForMultiCCS());
+    }
+}

shared/test/unit_test/release_helper/release_helper_tests_base.h

@@ -39,6 +39,8 @@ struct ReleaseHelperTestsBase : public ::testing::Test {
     void whenCallingAdjustMaxThreadsPerEuCountThenCorrectValueIsReturned();
     void whenShouldQueryPeerAccessCalledThenFalseReturned();
     void whenShouldQueryPeerAccessCalledThenTrueReturned();
+    void whenIsSingleDispatchRequiredForMultiCCSCalledThenFalseReturned();
+    void whenIsSingleDispatchRequiredForMultiCCSCalledThenTrueReturned();
     virtual std::vector<uint32_t> getRevisions() = 0;
 
     std::unique_ptr<ReleaseHelper> releaseHelper;