Use Eu per dss to callculate max work group size

Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>

Author: MaciejPlewka

opencl/source/kernel/kernel.cpp

@@ -77,7 +77,8 @@ Kernel::Kernel(Program *programArg, const KernelInfo &kernelInfoArg, ClDevice &c
     program->retainForKernel();
     imageTransformer.reset(new ImageTransformer);
     if (kernelInfoArg.kernelDescriptor.kernelAttributes.simdSize == 1u) {
-        maxKernelWorkGroupSize = HwHelper::get(getHardwareInfo().platform.eRenderCoreFamily).getMaxThreadsForWorkgroup(getHardwareInfo(), static_cast<uint32_t>(getDevice().getDevice().getDeviceInfo().maxNumEUsPerSubSlice));
+        auto deviceInfo = getDevice().getDevice().getDeviceInfo();
+        maxKernelWorkGroupSize = HwHelper::get(getHardwareInfo().platform.eRenderCoreFamily).getMaxThreadsForWorkgroupInDSSOrSS(getHardwareInfo(), static_cast<uint32_t>(deviceInfo.maxNumEUsPerSubSlice), static_cast<uint32_t>(deviceInfo.maxNumEUsPerDualSubSlice));
     } else {
         maxKernelWorkGroupSize = static_cast<uint32_t>(clDevice.getSharedDeviceInfo().maxWorkGroupSize);
     }

opencl/test/unit_test/device/device_caps_tests.cpp

@@ -1685,3 +1685,20 @@ HWTEST_F(QueueFamilyNameTest, givenBcsWhenGettingQueueFamilyNameThenReturnProper
 HWTEST_F(QueueFamilyNameTest, givenInvalidEngineGroupWhenGettingQueueFamilyNameThenReturnEmptyName) {
     verify(EngineGroupType::MaxEngineGroups, "");
 }
+HWCMDTEST_F(IGFX_GEN8_CORE, DeviceGetCapsTest, givenSysInfoWhenDeviceCreatedThenMaxWorkGroupCalculatedCorrectly) {
+    HardwareInfo myHwInfo = *defaultHwInfo;
+    GT_SYSTEM_INFO &mySysInfo = myHwInfo.gtSystemInfo;
+    PLATFORM &myPlatform = myHwInfo.platform;
+
+    mySysInfo.EUCount = 16;
+    mySysInfo.SubSliceCount = 4;
+    mySysInfo.DualSubSliceCount = 2;
+    mySysInfo.ThreadCount = 16 * 8;
+    myPlatform.usRevId = 0x4;
+    auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&myHwInfo));
+    auto minSimd = 8;
+
+    auto expectedWG = (mySysInfo.ThreadCount / mySysInfo.EUCount) * (mySysInfo.EUCount / mySysInfo.SubSliceCount) * minSimd;
+
+    EXPECT_EQ(expectedWG, device->sharedDeviceInfo.maxWorkGroupSize);
+}

opencl/test/unit_test/kernel/kernel_tests.cpp

@@ -2515,6 +2515,7 @@ HWTEST_F(KernelTest, givenKernelWhenDebugFlagToUseMaxSimdForCalculationsIsUsedTh
 
     mySysInfo.EUCount = 24;
     mySysInfo.SubSliceCount = 3;
+    mySysInfo.DualSubSliceCount = 3;
     mySysInfo.ThreadCount = 24 * 7;
     auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&myHwInfo));
 
@@ -3166,7 +3167,8 @@ TEST_F(KernelTests, givenKernelWithSimdEqual1WhenKernelCreatedThenMaxWorgGroupSi
     std::unique_ptr<MockKernel> pKernel(new MockKernel(pProgram, *pKernelInfo, *pClDevice));
 
     auto deviceMaxWorkGroupSize = pDevice->getDeviceInfo().maxWorkGroupSize;
-    auto maxThreadsPerWG = HwHelper::get(pKernel->getHardwareInfo().platform.eRenderCoreFamily).getMaxThreadsForWorkgroup(pKernel->getHardwareInfo(), static_cast<uint32_t>(pClDevice->getDevice().getDeviceInfo().maxNumEUsPerSubSlice));
+    auto deviceInfo = pClDevice->getDevice().getDeviceInfo();
+    auto maxThreadsPerWG = HwHelper::get(pKernel->getHardwareInfo().platform.eRenderCoreFamily).getMaxThreadsForWorkgroupInDSSOrSS(pKernel->getHardwareInfo(), static_cast<uint32_t>(deviceInfo.maxNumEUsPerSubSlice), static_cast<uint32_t>(deviceInfo.maxNumEUsPerDualSubSlice));
 
     EXPECT_LT(pKernel->getMaxKernelWorkGroupSize(), deviceMaxWorkGroupSize);
     EXPECT_EQ(pKernel->getMaxKernelWorkGroupSize(), maxThreadsPerWG);

shared/source/device/device_caps.cpp

@@ -107,9 +107,14 @@ void Device::initializeCaps() {
     deviceInfo.maxNumEUsPerSubSlice = (systemInfo.EuCountPerPoolMin == 0 || hwInfo.featureTable.ftrPooledEuEnabled == 0)
                                           ? (systemInfo.EUCount / systemInfo.SubSliceCount)
                                           : systemInfo.EuCountPerPoolMin;
+
+    deviceInfo.maxNumEUsPerDualSubSlice = (systemInfo.EuCountPerPoolMin == 0 || hwInfo.featureTable.ftrPooledEuEnabled == 0)
+                                              ? (systemInfo.EUCount / systemInfo.DualSubSliceCount)
+                                              : systemInfo.EuCountPerPoolMin;
+
     deviceInfo.numThreadsPerEU = systemInfo.ThreadCount / systemInfo.EUCount;
     deviceInfo.threadsPerEUConfigs = hwHelper.getThreadsPerEUConfigs();
-    auto maxWS = hwHelper.getMaxThreadsForWorkgroup(hwInfo, static_cast<uint32_t>(deviceInfo.maxNumEUsPerSubSlice)) * simdSizeUsed;
+    auto maxWS = hwHelper.getMaxThreadsForWorkgroupInDSSOrSS(hwInfo, static_cast<uint32_t>(deviceInfo.maxNumEUsPerSubSlice), static_cast<uint32_t>(deviceInfo.maxNumEUsPerDualSubSlice)) * simdSizeUsed;
 
     maxWS = Math::prevPowerOfTwo(maxWS);
     deviceInfo.maxWorkGroupSize = std::min(maxWS, 1024u);

shared/source/device/device_info.h

@@ -26,6 +26,7 @@ struct DeviceInfo {
     size_t imageMaxArraySize;
     size_t imageMaxBufferSize;
     size_t maxNumEUsPerSubSlice;
+    size_t maxNumEUsPerDualSubSlice;
     size_t maxParameterSize;
     size_t maxWorkGroupSize;
     size_t maxWorkItemSizes[3];

shared/source/helpers/hw_helper.h

@@ -97,6 +97,7 @@ class HwHelper {
     virtual std::string getExtensions() const = 0;
     static uint32_t getMaxThreadsForVfe(const HardwareInfo &hwInfo);
     virtual uint32_t getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const;
+    virtual uint32_t getMaxThreadsForWorkgroupInDSSOrSS(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice, uint32_t maxNumEUsPerDualSubSlice) const = 0;
     virtual uint32_t getMetricsLibraryGenId() const = 0;
     virtual uint32_t getMocsIndex(const GmmHelper &gmmHelper, bool l3enabled, bool l1enabled) const = 0;
     virtual bool tilingAllowed(bool isSharedContext, bool isImage1d, bool forceLinearStorage) = 0;
@@ -211,7 +212,7 @@ class HwHelperHw : public HwHelper {
 
     size_t getPaddingForISAAllocation() const override;
 
-    uint32_t getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const override;
+    uint32_t getMaxThreadsForWorkgroupInDSSOrSS(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice, uint32_t maxNumEUsPerDualSubSlice) const override;
 
     uint32_t getComputeUnitsUsedForScratch(const HardwareInfo *pHwInfo) const override;
 

shared/source/helpers/hw_helper_bdw_plus.inl

@@ -111,7 +111,7 @@ uint32_t HwHelperHw<GfxFamily>::getPlanarYuvMaxHeight() const {
 }
 
 template <typename GfxFamily>
-uint32_t HwHelperHw<GfxFamily>::getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const {
+uint32_t HwHelperHw<GfxFamily>::getMaxThreadsForWorkgroupInDSSOrSS(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice, uint32_t maxNumEUsPerDualSubSlice) const {
     return HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerSubSlice);
 }
 

shared/source/helpers/hw_helper_xehp_plus.inl

@@ -194,11 +194,11 @@ inline bool HwHelperHw<GfxFamily>::preferSmallWorkgroupSizeForKernel(const size_
 }
 
 template <typename GfxFamily>
-inline uint32_t HwHelperHw<GfxFamily>::getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const {
+inline uint32_t HwHelperHw<GfxFamily>::getMaxThreadsForWorkgroupInDSSOrSS(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice, uint32_t maxNumEUsPerDualSubSlice) const {
     if (isWorkaroundRequired(REVISION_A0, REVISION_B, hwInfo)) {
-        return std::min(HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerSubSlice), 64u);
+        return std::min(HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerDualSubSlice), 64u);
     }
-    return HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerSubSlice);
+    return HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerDualSubSlice);
 }
 
 } // namespace NEO