Add new function to set max threads for workgroup

Resolves: NEO-4116

Change-Id: I8b9faf582c42edcb6f616a2f4662200d0d5b73d3
Signed-off-by: Katarzyna Cencelewska <[email protected]>

Author: kcencele

core/helpers/hw_helper.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2017-2019 Intel Corporation
+ * Copyright (C) 2017-2020 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,4 +45,8 @@ uint32_t HwHelper::getMaxThreadsForVfe(const HardwareInfo &hwInfo) {
     return hwInfo.gtSystemInfo.EUCount * threadsPerEU;
 }
 
+uint32_t HwHelper::getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const {
+    uint32_t numThreadsPerEU = hwInfo.gtSystemInfo.ThreadCount / hwInfo.gtSystemInfo.EUCount;
+    return maxNumEUsPerSubSlice * numThreadsPerEU;
+}
 } // namespace NEO

core/helpers/hw_helper.h

@@ -62,6 +62,7 @@ class HwHelper {
     virtual bool getEnableLocalMemory(const HardwareInfo &hwInfo) const = 0;
     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 getMetricsLibraryGenId() const = 0;
     virtual uint32_t getMocsIndex(GmmHelper &gmmHelper, bool l3enabled, bool l1enabled) const = 0;
     virtual bool requiresAuxResolves() const = 0;
@@ -119,6 +120,8 @@ class HwHelperHw : public HwHelper {
 
     size_t getMaxBarrierRegisterPerSlice() const override;
 
+    uint32_t getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const override;
+
     uint32_t getComputeUnitsUsedForScratch(const HardwareInfo *pHwInfo) const override;
 
     uint32_t getPitchAlignmentForImage(const HardwareInfo *hwInfo) override;

core/helpers/hw_helper_base.inl

@@ -277,4 +277,8 @@ bool HwHelperHw<GfxFamily>::isForceEmuInt32DivRemSPWARequired(const HardwareInfo
     return false;
 }
 
+template <typename GfxFamily>
+uint32_t HwHelperHw<GfxFamily>::getMaxThreadsForWorkgroup(const HardwareInfo &hwInfo, uint32_t maxNumEUsPerSubSlice) const {
+    return HwHelper::getMaxThreadsForWorkgroup(hwInfo, maxNumEUsPerSubSlice);
+}
 } // namespace NEO

runtime/device/device_caps.cpp

@@ -277,16 +277,16 @@ void Device::initializeCaps() {
     deviceInfo.maxNumEUsPerSubSlice = 0;
     deviceInfo.maxSliceCount = systemInfo.SliceCount;
     deviceInfo.numThreadsPerEU = 0;
-    auto simdSizeUsed = DebugManager.flags.UseMaxSimdSizeToDeduceMaxWorkgroupSize.get() ? 32 : 8;
+    auto simdSizeUsed = DebugManager.flags.UseMaxSimdSizeToDeduceMaxWorkgroupSize.get() ? 32u : 8u;
 
     deviceInfo.maxNumEUsPerSubSlice = (systemInfo.EuCountPerPoolMin == 0 || hwInfo.featureTable.ftrPooledEuEnabled == 0)
                                           ? (systemInfo.EUCount / systemInfo.SubSliceCount)
                                           : systemInfo.EuCountPerPoolMin;
     deviceInfo.numThreadsPerEU = systemInfo.ThreadCount / systemInfo.EUCount;
-    auto maxWS = deviceInfo.maxNumEUsPerSubSlice * deviceInfo.numThreadsPerEU * simdSizeUsed;
+    auto maxWS = hwHelper.getMaxThreadsForWorkgroup(hwInfo, static_cast<uint32_t>(deviceInfo.maxNumEUsPerSubSlice)) * simdSizeUsed;
 
-    maxWS = Math::prevPowerOfTwo(uint32_t(maxWS));
-    deviceInfo.maxWorkGroupSize = std::min(uint32_t(maxWS), 1024u);
+    maxWS = Math::prevPowerOfTwo(maxWS);
+    deviceInfo.maxWorkGroupSize = std::min(maxWS, 1024u);
 
     // calculate a maximum number of subgroups in a workgroup (for the required SIMD size)
     deviceInfo.maxNumOfSubGroups = static_cast<uint32_t>(deviceInfo.maxWorkGroupSize / simdSizeUsed);

unit_tests/device/device_caps_tests.cpp

@@ -851,7 +851,7 @@ TEST(DeviceGetCaps, givenDisabledFtrPooledEuWhenCalculatingMaxEuPerSSThenIgnoreE
     EXPECT_EQ(expectedMaxWGS, device->getDeviceInfo().maxWorkGroupSize);
 }
 
-TEST(DeviceGetCaps, givenEnabledFtrPooledEuWhenCalculatingMaxEuPerSSThenDontIgnoreEuCountPerPoolMin) {
+HWTEST_F(DeviceGetCapsTest, givenEnabledFtrPooledEuWhenCalculatingMaxEuPerSSThenDontIgnoreEuCountPerPoolMin) {
     HardwareInfo myHwInfo = *platformDevices[0];
     GT_SYSTEM_INFO &mySysInfo = myHwInfo.gtSystemInfo;
     FeatureTable &mySkuTable = myHwInfo.featureTable;
@@ -884,7 +884,7 @@ TEST(DeviceGetCaps, givenDebugFlagToUseMaxSimdSizeForWkgCalculationWhenDeviceCap
     EXPECT_EQ(device->getDeviceInfo().maxWorkGroupSize / 32, device->getDeviceInfo().maxNumOfSubGroups);
 }
 
-TEST(DeviceGetCaps, givenDeviceThatHasHighNumberOfExecutionUnitsWhenMaxWorkgroupSizeIsComputedItIsLimitedTo1024) {
+HWTEST_F(DeviceGetCapsTest, givenDeviceThatHasHighNumberOfExecutionUnitsWhenMaxWorkgroupSizeIsComputedItIsLimitedTo1024) {
     HardwareInfo myHwInfo = *platformDevices[0];
     GT_SYSTEM_INFO &mySysInfo = myHwInfo.gtSystemInfo;