Disable cl_khr_subgroups extension on gen12.

Resolves: NEO-3622

Change-Id: Idab55e5fb6a51ea0a08714f447bb9fce85f1dbcd
Signed-off-by: Piotr Zdunowski <[email protected]>

Author: PiotrZdunowski

opencl/source/cl_device/cl_device_caps.cpp

@@ -116,9 +116,14 @@ void ClDevice::initializeCaps() {
     auto supportsVme = hwInfo.capabilityTable.supportsVme;
     auto supportsAdvancedVme = hwInfo.capabilityTable.supportsVme;
 
+    deviceInfo.independentForwardProgress = false;
+
     if (enabledClVersion >= 21) {
-        deviceInfo.independentForwardProgress = true;
-        deviceExtensions += "cl_khr_subgroups ";
+        if (hwHelper.isIndependentForwardProgressSupported()) {
+            deviceInfo.independentForwardProgress = true;
+            deviceExtensions += "cl_khr_subgroups ";
+        }
+
         deviceExtensions += "cl_khr_il_program ";
         if (supportsVme) {
             deviceExtensions += "cl_intel_spirv_device_side_avc_motion_estimation ";
@@ -128,8 +133,6 @@ void ClDevice::initializeCaps() {
         }
         deviceExtensions += "cl_intel_spirv_subgroups ";
         deviceExtensions += "cl_khr_spirv_no_integer_wrap_decoration ";
-    } else {
-        deviceInfo.independentForwardProgress = false;
     }
 
     if (enabledClVersion >= 20) {

opencl/source/gen12lp/hw_helper_gen12lp.cpp

@@ -150,6 +150,11 @@ std::string HwHelperHw<Family>::getExtensions() const {
     return "cl_intel_subgroup_local_block_io ";
 }
 
+template <>
+bool HwHelperHw<Family>::isIndependentForwardProgressSupported() {
+    return false;
+}
+
 template <>
 void MemorySynchronizationCommands<Family>::setExtraCacheFlushFields(Family::PIPE_CONTROL *pipeControl) {
     pipeControl->setHdcPipelineFlush(true);

opencl/test/unit_test/device/device_caps_tests.cpp

@@ -140,12 +140,6 @@ TEST_F(DeviceGetCapsTest, WhenCreatingDeviceThenCapsArePopulatedCorrectly) {
         EXPECT_EQ(expectedDeviceSubgroups[i], sharedCaps.maxSubGroups[i]);
     }
 
-    if (device->getEnabledClVersion() >= 21) {
-        EXPECT_TRUE(caps.independentForwardProgress != 0);
-    } else {
-        EXPECT_FALSE(caps.independentForwardProgress != 0);
-    }
-
     EXPECT_EQ(sharedCaps.maxWorkGroupSize / hwHelper.getMinimalSIMDSize(), caps.maxNumOfSubGroups);
 
     EXPECT_EQ(1024u, caps.maxOnDeviceEvents);
@@ -396,15 +390,6 @@ TEST_F(DeviceGetCapsTest, givenEnableSharingFormatQuerySetTrueAndEnabledMultiple
     EXPECT_THAT(caps.deviceExtensions, ::testing::Not(::testing::HasSubstr(std::string("cl_intel_sharing_format_query "))));
 }
 
-TEST_F(DeviceGetCapsTest, givenOpenCLVersion21WhenCapsAreCreatedThenDeviceReportsClKhrSubgroupsExtension) {
-    DebugManagerStateRestore dbgRestorer;
-    DebugManager.flags.ForceOCLVersion.set(21);
-    auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(defaultHwInfo.get()));
-    const auto &caps = device->getDeviceInfo();
-
-    EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_khr_subgroups")));
-}
-
 TEST_F(DeviceGetCapsTest, givenOpenCLVersion20WhenCapsAreCreatedThenDeviceDoesntReportClKhrSubgroupsExtension) {
     DebugManagerStateRestore dbgRestorer;
     DebugManager.flags.ForceOCLVersion.set(20);

opencl/test/unit_test/gen11/test_device_caps_gen11.cpp

@@ -70,3 +70,23 @@ GEN11TEST_F(Gen11DeviceCaps, givenGen11WhenCheckExtensionsThenSubgroupLocalBlock
 
     EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_intel_subgroup_local_block_io")));
 }
+
+GEN11TEST_F(Gen11DeviceCaps, givenGen11WhenCheckExtensionsThenDeviceProperlyReportsClKhrSubgroupsExtension) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_khr_subgroups")));
+    } else {
+        EXPECT_THAT(caps.deviceExtensions, ::testing::Not(testing::HasSubstr(std::string("cl_khr_subgroups"))));
+    }
+}
+
+GEN11TEST_F(Gen11DeviceCaps, givenGen11WhenCheckingCapsThenDeviceDoesProperlyReportsIndependentForwardProgress) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_TRUE(caps.independentForwardProgress != 0);
+    } else {
+        EXPECT_FALSE(caps.independentForwardProgress != 0);
+    }
+}

opencl/test/unit_test/gen12lp/test_device_caps_gen12lp.inl

@@ -34,6 +34,18 @@ TGLLPTEST_F(Gen12LpDeviceCaps, givenGen12lpWhenCheckExtensionsThenSubgroupLocalB
     EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_intel_subgroup_local_block_io")));
 }
 
+TGLLPTEST_F(Gen12LpDeviceCaps, givenGen12lpWhenCheckExtensionsThenDeviceDoesNotReportClKhrSubgroupsExtension) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    EXPECT_THAT(caps.deviceExtensions, ::testing::Not(testing::HasSubstr(std::string("cl_khr_subgroups"))));
+}
+
+TGLLPTEST_F(Gen12LpDeviceCaps, givenGen12lpWhenCheckingCapsThenDeviceDoesNotSupportIndependentForwardProgress) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    EXPECT_FALSE(caps.independentForwardProgress);
+}
+
 TGLLPTEST_F(Gen12LpDeviceCaps, allSkusSupportCorrectlyRoundedDivideSqrt) {
     const auto &caps = pClDevice->getDeviceInfo();
     EXPECT_EQ(0u, caps.singleFpConfig & CL_FP_CORRECTLY_ROUNDED_DIVIDE_SQRT);

opencl/test/unit_test/gen8/test_device_caps_gen8.cpp

@@ -18,6 +18,25 @@ GEN8TEST_F(Gen8DeviceCaps, defaultPreemptionMode) {
     EXPECT_TRUE(PreemptionMode::Disabled == pDevice->getHardwareInfo().capabilityTable.defaultPreemptionMode);
 }
 
+GEN8TEST_F(Gen8DeviceCaps, givenGen8WhenCheckExtensionsThenDeviceProperlyReportsClKhrSubgroupsExtension) {
+    const auto &caps = pClDevice->getDeviceInfo();
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_khr_subgroups")));
+    } else {
+        EXPECT_THAT(caps.deviceExtensions, ::testing::Not(testing::HasSubstr(std::string("cl_khr_subgroups"))));
+    }
+}
+
+GEN8TEST_F(Gen8DeviceCaps, givenGen8WhenCheckingCapsThenDeviceDoesProperlyReportsIndependentForwardProgress) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_TRUE(caps.independentForwardProgress != 0);
+    } else {
+        EXPECT_FALSE(caps.independentForwardProgress != 0);
+    }
+}
+
 GEN8TEST_F(Gen8DeviceCaps, kmdNotifyMechanism) {
     EXPECT_TRUE(pDevice->getHardwareInfo().capabilityTable.kmdNotifyProperties.enableKmdNotify);
     EXPECT_EQ(50000, pDevice->getHardwareInfo().capabilityTable.kmdNotifyProperties.delayKmdNotifyMicroseconds);

opencl/test/unit_test/gen9/test_device_caps_gen9.cpp

@@ -26,6 +26,25 @@ GEN9TEST_F(Gen9DeviceCaps, skuSpecificCaps) {
     }
 }
 
+GEN9TEST_F(Gen9DeviceCaps, givenGen9WhenCheckExtensionsThenDeviceProperlyReportsClKhrSubgroupsExtension) {
+    const auto &caps = pClDevice->getDeviceInfo();
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_THAT(caps.deviceExtensions, testing::HasSubstr(std::string("cl_khr_subgroups")));
+    } else {
+        EXPECT_THAT(caps.deviceExtensions, ::testing::Not(testing::HasSubstr(std::string("cl_khr_subgroups"))));
+    }
+}
+
+GEN9TEST_F(Gen9DeviceCaps, givenGen9WhenCheckingCapsThenDeviceDoesProperlyReportsIndependentForwardProgress) {
+    const auto &caps = pClDevice->getDeviceInfo();
+
+    if (pClDevice->getEnabledClVersion() >= 21) {
+        EXPECT_TRUE(caps.independentForwardProgress != 0);
+    } else {
+        EXPECT_FALSE(caps.independentForwardProgress != 0);
+    }
+}
+
 GEN9TEST_F(Gen9DeviceCaps, allSkusSupportCorrectlyRoundedDivideSqrt) {
     const auto &caps = pClDevice->getDeviceInfo();
     EXPECT_NE(0u, caps.singleFpConfig & CL_FP_CORRECTLY_ROUNDED_DIVIDE_SQRT);

opencl/test/unit_test/platform/platform_tests.cpp

@@ -96,8 +96,10 @@ TEST_F(PlatformTest, PlatformgetAsCompilerEnabledExtensionsString) {
     pPlatform->initializeWithNewDevices();
     auto compilerExtensions = pPlatform->getClDevice(0)->peekCompilerExtensions();
 
+    auto &hwHelper = HwHelper::get(pPlatform->getClDevice(0)->getHardwareInfo().platform.eRenderCoreFamily);
+
     EXPECT_THAT(compilerExtensions, ::testing::HasSubstr(std::string(" -cl-ext=-all,+cl")));
-    if (std::string(pPlatform->getClDevice(0)->getDeviceInfo().clVersion).find("OpenCL 2.1") != std::string::npos) {
+    if (std::string(pPlatform->getClDevice(0)->getDeviceInfo().clVersion).find("OpenCL 2.1") != std::string::npos && hwHelper.isIndependentForwardProgressSupported()) {
         EXPECT_THAT(compilerExtensions, ::testing::HasSubstr(std::string("cl_khr_subgroups")));
     }
 }

shared/source/helpers/hw_helper.h

@@ -83,6 +83,7 @@ class HwHelper {
     virtual bool isOffsetToSkipSetFFIDGPWARequired(const HardwareInfo &hwInfo) const = 0;
     virtual bool is3DPipelineSelectWARequired(const HardwareInfo &hwInfo) const = 0;
     virtual bool isFusedEuDispatchEnabled(const HardwareInfo &hwInfo) const = 0;
+    virtual bool isIndependentForwardProgressSupported() = 0;
 
     static uint32_t getSubDevicesCount(const HardwareInfo *pHwInfo);
     static uint32_t getEnginesCount(const HardwareInfo &hwInfo);
@@ -214,6 +215,8 @@ class HwHelperHw : public HwHelper {
 
     uint32_t getMinimalSIMDSize() override;
 
+    bool isIndependentForwardProgressSupported() override;
+
   protected:
     static const AuxTranslationMode defaultAuxTranslationMode;
     HwHelperHw() = default;

shared/source/helpers/hw_helper_bdw_plus.inl

@@ -66,6 +66,11 @@ uint32_t HwHelperHw<GfxFamily>::calculateAvailableThreadCount(PRODUCT_FAMILY fam
     return threadsPerEu * euCount;
 }
 
+template <typename GfxFamily>
+bool HwHelperHw<GfxFamily>::isIndependentForwardProgressSupported() {
+    return true;
+}
+
 template <typename GfxFamily>
 void MemorySynchronizationCommands<GfxFamily>::addPipeControlWA(LinearStream &commandStream, uint64_t gpuAddress, const HardwareInfo &hwInfo) {
 }