Change order of command queue groups

DziubanMaciejIntel · Compute-Runtime-Automation · commit eace896ec848 · 2021-09-16T17:39:52.000+02:00
Resolves: NEO-6257
Signed-off-by: Maciej Dziuban &lt;maciej.dziuban@intel.com&gt;
diff --git a/level_zero/core/test/unit_tests/gen12lp/test_device_gen12lp.cpp b/level_zero/core/test/unit_tests/gen12lp/test_device_gen12lp.cpp
@@ -211,14 +211,14 @@ HWTEST2_F(DeviceQueueGroupTest,
     EXPECT_EQ(ZE_RESULT_SUCCESS, res);
 
     for (uint32_t i = 0; i < count; i++) {
-        if (i == static_cast<uint32_t>(NEO::EngineGroupType::RenderCompute)) {
+        if (i == neoMockDevice->getIndexOfNonEmptyEngineGroup(NEO::EngineGroupType::RenderCompute)) {
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COMPUTE);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COPY);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COOPERATIVE_KERNELS);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_METRICS);
             EXPECT_EQ(properties[i].maxMemoryFillPatternSize, std::numeric_limits<size_t>::max());
             EXPECT_EQ(properties[i].numQueues, 1u);
-        } else if (i == static_cast<uint32_t>(NEO::EngineGroupType::Copy)) {
+        } else if (i == neoMockDevice->getIndexOfNonEmptyEngineGroup(NEO::EngineGroupType::Copy)) {
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COPY);
             EXPECT_EQ(properties[i].maxMemoryFillPatternSize, 4 * sizeof(uint32_t));
             EXPECT_EQ(properties[i].numQueues, 1u);
@@ -253,14 +253,14 @@ HWTEST2_F(DeviceQueueGroupTest,
     L0::Context *context = Context::fromHandle(hContext);
 
     for (uint32_t i = 0; i < count; i++) {
-        if (i == static_cast<uint32_t>(NEO::EngineGroupType::RenderCompute)) {
+        if (i == neoMockDevice->getIndexOfNonEmptyEngineGroup(NEO::EngineGroupType::RenderCompute)) {
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COMPUTE);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COPY);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COOPERATIVE_KERNELS);
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_METRICS);
             EXPECT_EQ(properties[i].maxMemoryFillPatternSize, std::numeric_limits<size_t>::max());
             EXPECT_EQ(properties[i].numQueues, 1u);
-        } else if (i == static_cast<uint32_t>(NEO::EngineGroupType::Copy)) {
+        } else if (i == neoMockDevice->getIndexOfNonEmptyEngineGroup(NEO::EngineGroupType::Copy)) {
             EXPECT_TRUE(properties[i].flags & ZE_COMMAND_QUEUE_GROUP_PROPERTY_FLAG_COPY);
             EXPECT_EQ(properties[i].maxMemoryFillPatternSize, 4 * sizeof(uint32_t));
             EXPECT_EQ(properties[i].numQueues, 1u);
diff --git a/level_zero/core/test/unit_tests/gen9/test_cmdqueue_gen9.cpp b/level_zero/core/test/unit_tests/gen9/test_cmdqueue_gen9.cpp
@@ -211,7 +211,8 @@ HWTEST2_F(CommandQueueGroupMultiDevice,
 
     L0::CommandQueueImp *cmdQueue = reinterpret_cast<CommandQueueImp *>(commandList0->cmdQImmediate);
     L0::DeviceImp *deviceImp = reinterpret_cast<L0::DeviceImp *>(device);
-    auto expectedCSR = deviceImp->neoDevice->getNearestGenericSubDevice(0)->getEngineGroups()[queueGroupOrdinal][queueGroupIndex].commandStreamReceiver;
+    const auto rcsIndex = static_cast<size_t>(NEO::EngineGroupType::RenderCompute);
+    auto expectedCSR = deviceImp->neoDevice->getNearestGenericSubDevice(0)->getEngineGroups()[rcsIndex][queueGroupIndex].commandStreamReceiver;
     EXPECT_EQ(cmdQueue->getCsr(), expectedCSR);
 }
 
diff --git a/opencl/test/unit_test/command_queue/command_queue_tests.cpp b/opencl/test/unit_test/command_queue/command_queue_tests.cpp
@@ -1677,7 +1677,7 @@ HWTEST_F(CommandQueueOnSpecificEngineTests, givenBcsFamilySelectedWhenCreatingQu
     EXPECT_EQ(properties[3], queueBcs.getQueueIndexWithinFamily());
 }
 
-HWTEST_F(CommandQueueOnSpecificEngineTests, givenNotInitializedCcsOsContextWhenCreatingQueueThenInitializeOsContext) {
+HWTEST_F(CommandQueueOnSpecificEngineTests, givenNotInitializedRcsOsContextWhenCreatingQueueThenInitializeOsContext) {
     DebugManagerStateRestore restore{};
     DebugManager.flags.NodeOrdinal.set(static_cast<int32_t>(aub_stream::EngineType::ENGINE_RCS));
     DebugManager.flags.DeferOsContextInitialization.set(1);
@@ -1689,7 +1689,27 @@ HWTEST_F(CommandQueueOnSpecificEngineTests, givenNotInitializedCcsOsContextWhenC
     OsContext &osContext = *context.getDevice(0)->getEngine(aub_stream::ENGINE_CCS, EngineUsage::Regular).osContext;
     EXPECT_FALSE(osContext.isInitialized());
 
-    fillProperties(properties, 1, 0);
+    const auto ccsFamilyIndex = static_cast<cl_uint>(context.getDevice(0)->getDevice().getIndexOfNonEmptyEngineGroup(EngineGroupType::Compute));
+    fillProperties(properties, ccsFamilyIndex, 0);
+    MockCommandQueueHw<FamilyType> queue(&context, context.getDevice(0), properties);
+    ASSERT_EQ(&osContext, queue.gpgpuEngine->osContext);
+    EXPECT_TRUE(osContext.isInitialized());
+}
+
+HWTEST_F(CommandQueueOnSpecificEngineTests, givenNotInitializedCcsOsContextWhenCreatingQueueThenInitializeOsContext) {
+    DebugManagerStateRestore restore{};
+    DebugManager.flags.NodeOrdinal.set(static_cast<int32_t>(aub_stream::EngineType::ENGINE_CCS));
+    DebugManager.flags.DeferOsContextInitialization.set(1);
+
+    auto raiiHwHelper = overrideHwHelper<FamilyType, MockHwHelper<FamilyType, 1, 1, 1>>();
+    MockContext context{};
+    cl_command_queue_properties properties[5] = {};
+
+    OsContext &osContext = *context.getDevice(0)->getEngine(aub_stream::ENGINE_RCS, EngineUsage::Regular).osContext;
+    EXPECT_FALSE(osContext.isInitialized());
+
+    const auto rcsFamilyIndex = static_cast<cl_uint>(context.getDevice(0)->getDevice().getIndexOfNonEmptyEngineGroup(EngineGroupType::RenderCompute));
+    fillProperties(properties, rcsFamilyIndex, 0);
     MockCommandQueueHw<FamilyType> queue(&context, context.getDevice(0), properties);
     ASSERT_EQ(&osContext, queue.gpgpuEngine->osContext);
     EXPECT_TRUE(osContext.isInitialized());
diff --git a/opencl/test/unit_test/command_queue/get_command_queue_info_tests.cpp b/opencl/test/unit_test/command_queue/get_command_queue_info_tests.cpp
@@ -219,7 +219,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, GetCommandQueueFamilyInfoTests, givenFamilyIdWhenGe
     clReleaseContext(context);
 }
 
-HWCMDTEST_F(IGFX_XE_HP_CORE, GetCommandQueueFamilyInfoTests, givenFamilyIdWhenCreatingCommandQueueForRootDeviceWithMultipleSubDevicesThenInvalidValueIsReturned) {
+HWCMDTEST_F(IGFX_XE_HP_CORE, GetCommandQueueFamilyInfoTests, givenNonZeroFamilyIdWhenCreatingCommandQueueForRootDeviceWithMultipleSubDevicesThenInvalidValueIsReturned) {
     DebugManagerStateRestore restorer;
     DebugManager.flags.CreateMultipleSubDevices.set(2);
     initPlatform();
@@ -228,7 +228,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, GetCommandQueueFamilyInfoTests, givenFamilyIdWhenCr
     const cl_device_id deviceId = rootDevice;
     auto context = clCreateContext(nullptr, 1, &deviceId, nullptr, nullptr, nullptr);
 
-    cl_command_queue_properties properties[] = {CL_QUEUE_FAMILY_INTEL, static_cast<uint32_t>(EngineGroupType::Compute), CL_QUEUE_INDEX_INTEL, 0, 0};
+    cl_command_queue_properties properties[] = {CL_QUEUE_FAMILY_INTEL, 1u, CL_QUEUE_INDEX_INTEL, 0, 0};
     EXPECT_EQ(2u, rootDevice->getNumGenericSubDevices());
     cl_int retVal;
     auto commandQueue = clCreateCommandQueueWithProperties(context, rootDevice, properties, &retVal);
@@ -266,4 +266,4 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, MultiEngineQueueHwTests, givenLimitedNumberOfCcsWhe
         EXPECT_EQ(CL_SUCCESS, retVal);
         clReleaseCommandQueue(clCommandQueue);
     }
-}
+}
diff --git a/opencl/test/unit_test/device/device_tests.cpp b/opencl/test/unit_test/device/device_tests.cpp
@@ -11,6 +11,7 @@
 #include "shared/source/indirect_heap/indirect_heap.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
+#include "shared/test/common/helpers/engine_descriptor_helper.h"
 #include "shared/test/common/helpers/ult_hw_config.h"
 #include "shared/test/common/helpers/unit_test_helper.h"
 #include "shared/test/common/helpers/variable_backup.h"
@@ -21,6 +22,7 @@
 
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
+#include "opencl/test/unit_test/helpers/raii_hw_helper.h"
 #include "opencl/test/unit_test/libult/ult_command_stream_receiver.h"
 #include "opencl/test/unit_test/mocks/mock_context.h"
 #include "opencl/test/unit_test/mocks/mock_csr.h"
@@ -557,6 +559,36 @@ HWTEST_F(DeviceHwTest, givenDeviceCreationWhenCsrFailsToCreateGlobalSyncAllocati
     EXPECT_EQ(nullptr, mockDevice);
 }
 
+HWTEST_F(DeviceHwTest, givenBothCcsAndRcsEnginesInDeviceWhenGettingFirstEngineGroupsThenReturnCcs) {
+    struct MyHwHelper : HwHelperHw<FamilyType> {
+        EngineGroupType getEngineGroupType(aub_stream::EngineType engineType, EngineUsage engineUsage, const HardwareInfo &hwInfo) const {
+            if (engineType == aub_stream::ENGINE_RCS) {
+                return EngineGroupType::RenderCompute;
+            }
+            if (EngineHelpers::isCcs(engineType)) {
+                return EngineGroupType::Compute;
+            }
+            UNRECOVERABLE_IF(true);
+        }
+    };
+    RAIIHwHelperFactory<MyHwHelper> overrideHwHelper{::defaultHwInfo->platform.eRenderCoreFamily};
+
+    MockOsContext rcsContext(0, EngineDescriptorHelper::getDefaultDescriptor({aub_stream::EngineType::ENGINE_RCS, EngineUsage::Regular}));
+    EngineControl rcsEngine{nullptr, &rcsContext};
+
+    MockOsContext ccsContext(1, EngineDescriptorHelper::getDefaultDescriptor({aub_stream::EngineType::ENGINE_CCS, EngineUsage::Regular}));
+    EngineControl ccsEngine{nullptr, &ccsContext};
+
+    MockDevice device{};
+    ASSERT_EQ(nullptr, device.getNonEmptyEngineGroup(0u));
+    device.addEngineToEngineGroup(rcsEngine);
+    device.addEngineToEngineGroup(ccsEngine);
+
+    const auto firstGroup = device.getNonEmptyEngineGroup(0u);
+    EXPECT_EQ(1u, firstGroup->size());
+    EXPECT_EQ(aub_stream::EngineType::ENGINE_CCS, firstGroup->at(0).getEngineType());
+}
+
 TEST(DeviceGetEngineTest, givenHwCsrModeWhenGetEngineThenDedicatedForInternalUsageEngineIsReturned) {
     auto device = std::unique_ptr<Device>(MockDevice::createWithNewExecutionEnvironment<Device>(nullptr));
 
diff --git a/shared/source/helpers/definitions/engine_group_types.h b/shared/source/helpers/definitions/engine_group_types.h
@@ -12,8 +12,8 @@
 namespace NEO {
 
 enum class EngineGroupType : uint32_t {
-    RenderCompute = 0,
-    Compute,
+    Compute = 0,
+    RenderCompute,
     Copy,
     MaxEngineGroups
 };
diff --git a/shared/test/common/mocks/mock_device.h b/shared/test/common/mocks/mock_device.h
@@ -46,6 +46,7 @@ struct MockSubDevice : public SubDevice {
 
 class MockDevice : public RootDevice {
   public:
+    using Device::addEngineToEngineGroup;
     using Device::commandStreamReceivers;
     using Device::createDeviceInternals;
     using Device::createEngine;

Original file line number	Diff line number	Diff line change
`@@ -211,7 +211,8 @@ HWTEST2_F(CommandQueueGroupMultiDevice,`
`211`	`211`
`212`	`212`	`L0::CommandQueueImp cmdQueue = reinterpret_cast<CommandQueueImp >(commandList0->cmdQImmediate);`
`213`	`213`	`L0::DeviceImp deviceImp = reinterpret_cast<L0::DeviceImp >(device);`
`214`		`- auto expectedCSR = deviceImp->neoDevice->getNearestGenericSubDevice(0)->getEngineGroups()[queueGroupOrdinal][queueGroupIndex].commandStreamReceiver;`
	`214`	`+ const auto rcsIndex = static_cast<size_t>(NEO::EngineGroupType::RenderCompute);`
	`215`	`+ auto expectedCSR = deviceImp->neoDevice->getNearestGenericSubDevice(0)->getEngineGroups()[rcsIndex][queueGroupIndex].commandStreamReceiver;`
`215`	`216`	`EXPECT_EQ(cmdQueue->getCsr(), expectedCSR);`
`216`	`217`	`}`
`217`	`218`