intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-19 06:24:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

test: correct expectations in shared ults if heapless enabled

Browse Source 
Related-To: NEO-10641
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
This commit is contained in:
Kamil Kopryk
2024-08-26 08:40:34 +00:00
committed by Compute-Runtime-Automation
parent df9fa343d1
commit fc3646b58c
12 changed files with 223 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
shared/test/common/fixtures/memory_allocator_fixture.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2023 Intel Corporation
* Copyright (C) 2018-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -25,6 +25,9 @@ class MemoryAllocatorFixture : public MemoryManagementFixture {
executionEnvironment->prepareRootDeviceEnvironments(1);
executionEnvironment->rootDeviceEnvironments[0]->setHwInfoAndInitHelpers(defaultHwInfo.get());
executionEnvironment->rootDeviceEnvironments[0]->initGmm();
UnitTestSetter::setRcsExposure(*executionEnvironment->rootDeviceEnvironments[0]);
UnitTestSetter::setCcsExposure(*executionEnvironment->rootDeviceEnvironments[0]);
executionEnvironment->calculateMaxOsContextCount();
device.reset(MockDevice::createWithExecutionEnvironment<MockDevice>(defaultHwInfo.get(), executionEnvironment, 0u));
memoryManager = new MockMemoryManager(false, false, *executionEnvironment);

1
shared/test/common/helpers/unit_test_helper.h
View File

@@ -119,6 +119,7 @@ struct UnitTestHelper {
static GenCmdList::iterator findWalkerTypeCmd(GenCmdList::iterator begin, GenCmdList::iterator end);
static std::vector<GenCmdList::iterator> findAllWalkerTypeCmds(GenCmdList::iterator begin, GenCmdList::iterator end);
static typename GfxFamily::WalkerVariant getWalkerVariant(void *walkerItor);
static void getSpaceAndInitWalkerCmd(LinearStream &stream, bool heapless);
static size_t getWalkerSize(bool isHeaplessEnabled);
};

6
shared/test/common/helpers/unit_test_helper_bdw_and_later.inl
View File

@@ -169,4 +169,10 @@ uint64_t UnitTestHelper<GfxFamily>::getWalkerPartitionEstimateSpaceRequiredInCom
return 0u;
}
template <typename GfxFamily>
void UnitTestHelper<GfxFamily>::getSpaceAndInitWalkerCmd(LinearStream &stream, bool heapless) {
using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
*stream.getSpaceForCmd<GPGPU_WALKER>() = GfxFamily::template getInitGpuWalker<GPGPU_WALKER>();
}
} // namespace NEO

6
shared/test/common/helpers/unit_test_helper_xehp_and_later.inl
View File

@@ -156,4 +156,10 @@ uint64_t UnitTestHelper<GfxFamily>::getWalkerPartitionEstimateSpaceRequiredInCom
return WalkerPartition::estimateSpaceRequiredInCommandBuffer<GfxFamily, DefaultWalkerType>(testArgs);
}
template <typename GfxFamily>
void UnitTestHelper<GfxFamily>::getSpaceAndInitWalkerCmd(LinearStream &stream, bool heapless) {
using COMPUTE_WALKER = typename GfxFamily::COMPUTE_WALKER;
*stream.getSpaceForCmd<COMPUTE_WALKER>() = GfxFamily::template getInitGpuWalker<COMPUTE_WALKER>();
}
} // namespace NEO

1
shared/test/common/libult/ult_command_stream_receiver.h
View File

@@ -117,6 +117,7 @@ class UltCommandStreamReceiver : public CommandStreamReceiverHw<GfxFamily>, publ
using BaseClass::CommandStreamReceiver::flushBcsTask;
using BaseClass::CommandStreamReceiver::flushStamp;
using BaseClass::CommandStreamReceiver::globalFenceAllocation;
using BaseClass::CommandStreamReceiver::globalStatelessHeapAllocation;
using BaseClass::CommandStreamReceiver::gpuHangCheckPeriod;
using BaseClass::CommandStreamReceiver::gsbaFor32BitProgrammed;
using BaseClass::CommandStreamReceiver::heaplessModeEnabled;

13
shared/test/unit_test/built_ins/sip_tests.cpp
View File

@@ -723,15 +723,24 @@ TEST(DebugBindlessSip, givenOfflineDebuggingModeWhenDebugSipForContextIsCreatedT
executionEnvironment->setDebuggingMode(DebuggingMode::offline);
auto mockDevice = std::unique_ptr<MockDevice>(MockDevice::createWithExecutionEnvironment<MockDevice>(defaultHwInfo.get(), executionEnvironment, 0u));
auto osContext = std::make_unique<OsContextMock>(0, EngineDescriptorHelper::getDefaultDescriptor({aub_stream::ENGINE_RCS, EngineUsage::regular}));
auto osContext = std::make_unique<OsContextMock>(0, EngineDescriptorHelper::getDefaultDescriptor({aub_stream::ENGINE_CCS, EngineUsage::regular}));
osContext->debuggableContext = true;
osContext->offlineDumpCtxId = uint64_t(0xAA) << 32 | 0xBB;
auto csr = mockDevice->createCommandStreamReceiver();
csr->setupContext(*osContext);
EXPECT_NE(nullptr, mockDevice);
// ensure once_flag has not already been satisfied
auto contextId = osContext->getContextId();
auto it = builtIns->perContextSipKernels.find(contextId);
if (it != builtIns->perContextSipKernels.end()) {
mockDevice->getMemoryManager()->freeGraphicsMemory(it->second.first->getSipAllocation());
builtIns->perContextSipKernels.erase(contextId);
}
builtIns->perContextSipKernels.erase(contextId);
auto &sipKernel = NEO::SipKernel::getBindlessDebugSipKernel(*mockDevice, &csr->getOsContext());
EXPECT_NE(nullptr, &sipKernel);

202
shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
View File

@@ -78,6 +78,10 @@ struct CommandStreamReceiverTest : public DeviceFixture,
ASSERT_NE(nullptr, commandStreamReceiver);
memoryManager = commandStreamReceiver->getMemoryManager();
internalAllocationStorage = commandStreamReceiver->getInternalAllocationStorage();
auto &compilerProductHelper = pDevice->getCompilerProductHelper();
auto heaplessEnabled = compilerProductHelper.isHeaplessModeEnabled();
this->heaplessStateInit = compilerProductHelper.isHeaplessStateInitEnabled(heaplessEnabled);
}
void TearDown() override {
@@ -87,6 +91,7 @@ struct CommandStreamReceiverTest : public DeviceFixture,
CommandStreamReceiver *commandStreamReceiver = nullptr;
MemoryManager *memoryManager = nullptr;
InternalAllocationStorage *internalAllocationStorage = nullptr;
bool heaplessStateInit = false;
};
TEST_F(CommandStreamReceiverTest, givenOsAgnosticCsrWhenGettingCompletionValueThenProperTaskCountIsReturned) {
@@ -118,8 +123,9 @@ TEST_F(CommandStreamReceiverTest, givenBaseCsrWhenCallingWaitUserFenceThenReturn
HWTEST_F(CommandStreamReceiverTest, WhenCreatingCsrThenDefaultValuesAreSet) {
auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
EXPECT_EQ(0u, csr.peekTaskLevel());
EXPECT_EQ(0u, csr.peekTaskCount());
EXPECT_EQ(csr.heaplessStateInitialized ? 1u : 0u, csr.peekTaskCount());
EXPECT_FALSE(csr.isPreambleSent);
}
@@ -553,7 +559,7 @@ HWTEST_F(CommandStreamReceiverTest, givenFlushUnsuccessWhenWaitingForTaskCountTh
auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
csr.activePartitions = 1;
constexpr auto taskCountToWait = 1;
auto taskCountToWait = csr.heaplessStateInitialized ? 2 : 1;
csr.flushReturnValue = SubmissionStatus::failed;
auto waitStatus = csr.waitForTaskCount(taskCountToWait);
EXPECT_EQ(WaitStatus::notReady, waitStatus);
@@ -1889,8 +1895,11 @@ TEST(CommandStreamReceiverMultiContextTests, givenMultipleCsrsWhenSameResourcesA
EXPECT_EQ(1u, commandStreamReceiver0.getResidencyAllocations().size());
EXPECT_EQ(1u, commandStreamReceiver1.getResidencyAllocations().size());
EXPECT_EQ(1u, graphicsAllocation.getResidencyTaskCount(csr0ContextId));
EXPECT_EQ(1u, graphicsAllocation.getResidencyTaskCount(csr1ContextId));
auto expectedTaskCount0 = commandStreamReceiver0.peekTaskCount() + 1;
auto expectedTaskCount1 = commandStreamReceiver1.peekTaskCount() + 1;
EXPECT_EQ(expectedTaskCount0, graphicsAllocation.getResidencyTaskCount(csr0ContextId));
EXPECT_EQ(expectedTaskCount1, graphicsAllocation.getResidencyTaskCount(csr1ContextId));
commandStreamReceiver0.makeNonResident(graphicsAllocation);
EXPECT_FALSE(graphicsAllocation.isResident(csr0ContextId));
@@ -2516,7 +2525,9 @@ TEST_F(CommandStreamReceiverTest, givenPreambleFlagIsSetWhenGettingFlagStateThen
TEST_F(CommandStreamReceiverTest, givenPreemptionSentIsInitialWhenSettingPreemptionToNewModeThenExpectCorrectPreemption) {
PreemptionMode mode = PreemptionMode::Initial;
EXPECT_EQ(mode, commandStreamReceiver->getPreemptionMode());
if (!heaplessStateInit) {
EXPECT_EQ(mode, commandStreamReceiver->getPreemptionMode());
}
mode = PreemptionMode::ThreadGroup;
commandStreamReceiver->setPreemptionMode(mode);
EXPECT_EQ(mode, commandStreamReceiver->getPreemptionMode());
@@ -3209,11 +3220,12 @@ HWTEST_F(CommandStreamReceiverHwTest, givenOutOfMemoryFailureOnFlushWhenFlushing
HWTEST_F(CommandStreamReceiverHwTest, givenOutOfMemoryFailureOnFlushWhenInitializingDeviceWithFirstSubmissionThenErrorIsPropagated) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
commandStreamReceiver.latestFlushedTaskCount = 0;
commandStreamReceiver.flushReturnValue = SubmissionStatus::outOfMemory;
EXPECT_EQ(SubmissionStatus::outOfMemory, commandStreamReceiver.initializeDeviceWithFirstSubmission(*pDevice));
commandStreamReceiver.latestFlushedTaskCount = 0;
commandStreamReceiver.flushReturnValue = SubmissionStatus::outOfHostMemory;
EXPECT_EQ(SubmissionStatus::outOfHostMemory, commandStreamReceiver.initializeDeviceWithFirstSubmission(*pDevice));
}
@@ -3376,7 +3388,11 @@ HWTEST_F(CommandStreamReceiverTest, givenL1CachePolicyInitializedInCsrWhenGettin
HWTEST_F(CommandStreamReceiverHwTest, givenCreateGlobalStatelessHeapAllocationWhenGettingIndirectHeapObjectThenHeapAndAllocationAreInitialized) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
EXPECT_EQ(nullptr, commandStreamReceiver.getGlobalStatelessHeap());
if (commandStreamReceiver.heaplessStateInitialized) {
EXPECT_NE(nullptr, commandStreamReceiver.getGlobalStatelessHeap());
} else {
EXPECT_EQ(nullptr, commandStreamReceiver.getGlobalStatelessHeap());
}
commandStreamReceiver.createGlobalStatelessHeap();
auto heap = commandStreamReceiver.getGlobalStatelessHeap();
@@ -3397,6 +3413,11 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.storeMakeResidentAllocations = true;
EXPECT_EQ(nullptr, commandStreamReceiver.getGlobalStatelessHeap());
@@ -3427,6 +3448,10 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using STATE_COMPUTE_MODE = typename FamilyType::STATE_COMPUTE_MODE;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getStateComputeModeDirty());
commandStreamReceiver.flushTask(commandStream,
@@ -3778,6 +3803,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using STATE_COMPUTE_MODE = typename FamilyType::STATE_COMPUTE_MODE;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getStateComputeModeDirty());
@@ -3808,6 +3836,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using STATE_COMPUTE_MODE = typename FamilyType::STATE_COMPUTE_MODE;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getStateComputeModeDirty());
@@ -3851,6 +3882,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getGSBAStateDirty());
@@ -3889,6 +3923,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getGSBAStateDirty());
@@ -3929,6 +3966,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
EXPECT_TRUE(commandStreamReceiver.getGSBAStateDirty());
@@ -3983,6 +4023,10 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.createGlobalStatelessHeap();
EXPECT_TRUE(commandStreamReceiver.getGSBAStateDirty());
@@ -4024,6 +4068,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using _3DSTATE_BINDING_TABLE_POOL_ALLOC = typename FamilyType::_3DSTATE_BINDING_TABLE_POOL_ALLOC;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.createGlobalStatelessHeap();
EXPECT_TRUE(commandStreamReceiver.getGSBAStateDirty());
@@ -4306,21 +4353,30 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using DefaultWalkerType = typename FamilyType::DefaultWalkerType;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
auto heapless = commandStreamReceiver.heaplessModeEnabled;
auto heaplessStateInit = commandStreamReceiver.heaplessStateInitialized;
bool additionalSyncCmd = NEO::MemorySynchronizationCommands<FamilyType>::getSizeForSingleAdditionalSynchronization(commandStreamReceiver.peekRootDeviceEnvironment()) > 0;
commandStreamReceiver.storeMakeResidentAllocations = true;
auto startOffset = commandStream.getUsed();
auto immediateListCmdBufferAllocation = commandStream.getGraphicsAllocation();
*commandStream.getSpaceForCmd<DefaultWalkerType>() = FamilyType::template getInitGpuWalker<DefaultWalkerType>();
UnitTestHelper<FamilyType>::getSpaceAndInitWalkerCmd(commandStream, heapless);
auto csrTagAllocation = commandStreamReceiver.getTagAllocation();
uint64_t postsyncAddress = csrTagAllocation->getGpuAddress();
immediateFlushTaskFlags.blockingAppend = true;
auto completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(1u, completionStamp.taskCount);
EXPECT_EQ(1u, commandStreamReceiver.taskCount);
EXPECT_EQ(1u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(1u, commandStreamReceiver.latestFlushedTaskCount);
auto completionStamp = heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, startOffset, immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
auto expectedCount = heaplessStateInit ? 2u : 1u;
EXPECT_EQ(expectedCount, completionStamp.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
HardwareParse hwParserCsr;
hwParserCsr.parseCommands<FamilyType>(commandStream, 0);
@@ -4329,7 +4385,7 @@ HWTEST2_F(CommandStreamReceiverHwTest,
auto pipeControlCmd = genCmdCast<PIPE_CONTROL *>(*cmdItor);
ASSERT_NE(nullptr, pipeControlCmd);
EXPECT_EQ(postsyncAddress, NEO::UnitTestHelper<FamilyType>::getPipeControlPostSyncAddress(*pipeControlCmd));
EXPECT_EQ(1u, pipeControlCmd->getImmediateData());
EXPECT_EQ(expectedCount, pipeControlCmd->getImmediateData());
cmdItor++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdItor);
@@ -4346,13 +4402,19 @@ HWTEST2_F(CommandStreamReceiverHwTest,
startOffset = commandStream.getUsed();
EXPECT_EQ(0u, (startOffset % MemoryConstants::cacheLineSize));
*commandStream.getSpaceForCmd<DefaultWalkerType>() = FamilyType::template getInitGpuWalker<DefaultWalkerType>();
UnitTestHelper<FamilyType>::getSpaceAndInitWalkerCmd(commandStream, heapless);
completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(2u, completionStamp.taskCount);
EXPECT_EQ(2u, commandStreamReceiver.taskCount);
EXPECT_EQ(2u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(2u, commandStreamReceiver.latestFlushedTaskCount);
if (heaplessStateInit) {
completionStamp = commandStreamReceiver.flushImmediateTaskStateless(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
} else {
completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
}
expectedCount++;
EXPECT_EQ(expectedCount, completionStamp.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
hwParserCsr.tearDown();
hwParserCsr.parseCommands<FamilyType>(commandStream, startOffset);
@@ -4361,7 +4423,7 @@ HWTEST2_F(CommandStreamReceiverHwTest,
pipeControlCmd = genCmdCast<PIPE_CONTROL *>(*cmdItor);
ASSERT_NE(nullptr, pipeControlCmd);
EXPECT_EQ(postsyncAddress, NEO::UnitTestHelper<FamilyType>::getPipeControlPostSyncAddress(*pipeControlCmd));
EXPECT_EQ(2u, pipeControlCmd->getImmediateData());
EXPECT_EQ(expectedCount, pipeControlCmd->getImmediateData());
cmdItor++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdItor);
@@ -4387,21 +4449,27 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using DefaultWalkerType = typename FamilyType::DefaultWalkerType;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
auto heapless = commandStreamReceiver.heaplessModeEnabled;
auto heaplessStateInit = commandStreamReceiver.heaplessStateInitialized;
bool additionalSyncCmd = NEO::MemorySynchronizationCommands<FamilyType>::getSizeForSingleAdditionalSynchronization(commandStreamReceiver.peekRootDeviceEnvironment()) > 0;
commandStreamReceiver.storeMakeResidentAllocations = true;
auto startOffset = commandStream.getUsed();
auto immediateListCmdBufferAllocation = commandStream.getGraphicsAllocation();
*commandStream.getSpaceForCmd<DefaultWalkerType>() = FamilyType::template getInitGpuWalker<DefaultWalkerType>();
UnitTestHelper<FamilyType>::getSpaceAndInitWalkerCmd(commandStream, heapless);
auto csrTagAllocation = commandStreamReceiver.getTagAllocation();
uint64_t postsyncAddress = csrTagAllocation->getGpuAddress();
immediateFlushTaskFlags.requireTaskCountUpdate = true;
auto completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(1u, completionStamp.taskCount);
EXPECT_EQ(1u, commandStreamReceiver.taskCount);
EXPECT_EQ(1u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(1u, commandStreamReceiver.latestFlushedTaskCount);
auto completionStamp = heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, startOffset, immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
auto expectedCount = heaplessStateInit ? 2u : 1u;
EXPECT_EQ(expectedCount, completionStamp.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
HardwareParse hwParserCsr;
hwParserCsr.parseCommands<FamilyType>(commandStream, 0);
@@ -4410,7 +4478,7 @@ HWTEST2_F(CommandStreamReceiverHwTest,
auto pipeControlCmd = genCmdCast<PIPE_CONTROL *>(*cmdItor);
ASSERT_NE(nullptr, pipeControlCmd);
EXPECT_EQ(postsyncAddress, NEO::UnitTestHelper<FamilyType>::getPipeControlPostSyncAddress(*pipeControlCmd));
EXPECT_EQ(1u, pipeControlCmd->getImmediateData());
EXPECT_EQ(expectedCount, pipeControlCmd->getImmediateData());
cmdItor++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdItor);
@@ -4427,13 +4495,15 @@ HWTEST2_F(CommandStreamReceiverHwTest,
startOffset = commandStream.getUsed();
EXPECT_EQ(0u, (startOffset % MemoryConstants::cacheLineSize));
*commandStream.getSpaceForCmd<DefaultWalkerType>() = FamilyType::template getInitGpuWalker<DefaultWalkerType>();
UnitTestHelper<FamilyType>::getSpaceAndInitWalkerCmd(commandStream, heapless);
expectedCount++;
completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(2u, completionStamp.taskCount);
EXPECT_EQ(2u, commandStreamReceiver.taskCount);
EXPECT_EQ(2u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(2u, commandStreamReceiver.latestFlushedTaskCount);
EXPECT_EQ(expectedCount, completionStamp.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
hwParserCsr.tearDown();
hwParserCsr.parseCommands<FamilyType>(commandStream, startOffset);
@@ -4442,7 +4512,7 @@ HWTEST2_F(CommandStreamReceiverHwTest,
pipeControlCmd = genCmdCast<PIPE_CONTROL *>(*cmdItor);
ASSERT_NE(nullptr, pipeControlCmd);
EXPECT_EQ(postsyncAddress, NEO::UnitTestHelper<FamilyType>::getPipeControlPostSyncAddress(*pipeControlCmd));
EXPECT_EQ(2u, pipeControlCmd->getImmediateData());
EXPECT_EQ(expectedCount, pipeControlCmd->getImmediateData());
cmdItor++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdItor);
@@ -4468,6 +4538,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
commandStreamReceiver.storeMakeResidentAllocations = true;
commandStreamReceiver.recordFlushedBatchBuffer = true;
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
auto startOffset = commandStream.getUsed();
auto immediateListCmdBufferAllocation = commandStream.getGraphicsAllocation();
@@ -4525,26 +4598,32 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using DefaultWalkerType = typename FamilyType::DefaultWalkerType;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
bool heapless = commandStreamReceiver.heaplessModeEnabled;
bool heaplessStateInit = commandStreamReceiver.heaplessStateInitialized;
commandStreamReceiver.storeMakeResidentAllocations = true;
commandStreamReceiver.recordFlushedBatchBuffer = true;
auto startOffset = commandStream.getUsed();
auto immediateListCmdBufferAllocation = commandStream.getGraphicsAllocation();
*commandStream.getSpaceForCmd<DefaultWalkerType>() = FamilyType::template getInitGpuWalker<DefaultWalkerType>();
UnitTestHelper<FamilyType>::getSpaceAndInitWalkerCmd(commandStream, heapless);
immediateFlushTaskFlags.blockingAppend = true;
commandStreamReceiver.flushReturnValue = NEO::SubmissionStatus::failed;
auto completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
auto completionStamp = heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, startOffset, immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
auto csrCmdBufferAllocation = commandStreamReceiver.commandStream.getGraphicsAllocation();
TaskCountType currentTaskCountType = 1u;
auto expectedCount = heaplessStateInit ? 1u : 0u;
EXPECT_EQ(NEO::CompletionStamp::failed, completionStamp.taskCount);
EXPECT_EQ(0u, commandStreamReceiver.taskCount);
EXPECT_EQ(0u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(0u, commandStreamReceiver.latestFlushedTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
EXPECT_FALSE(commandStreamReceiver.isMadeResident(csrCmdBufferAllocation, currentTaskCountType));
EXPECT_TRUE(commandStreamReceiver.isMadeResident(immediateListCmdBufferAllocation, currentTaskCountType));
@@ -4555,14 +4634,19 @@ HWTEST2_F(CommandStreamReceiverHwTest,
EXPECT_EQ(false, recordedBatchBuffer.hasStallingCmds);
EXPECT_EQ(false, recordedBatchBuffer.hasRelaxedOrderingDependencies);
completionStamp = commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
completionStamp = heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, startOffset, immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(NEO::CompletionStamp::failed, completionStamp.taskCount);
EXPECT_EQ(0u, commandStreamReceiver.taskCount);
EXPECT_EQ(0u, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(0u, commandStreamReceiver.latestFlushedTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.taskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestSentTaskCount);
EXPECT_EQ(expectedCount, commandStreamReceiver.latestFlushedTaskCount);
EXPECT_FALSE(commandStreamReceiver.isMadeResident(immediateListCmdBufferAllocation, currentTaskCountType));
if (heaplessStateInit) {
EXPECT_TRUE(commandStreamReceiver.isMadeResident(immediateListCmdBufferAllocation, currentTaskCountType));
} else {
EXPECT_FALSE(commandStreamReceiver.isMadeResident(immediateListCmdBufferAllocation, currentTaskCountType));
}
}
HWTEST2_F(CommandStreamReceiverHwTest,
@@ -4571,12 +4655,18 @@ HWTEST2_F(CommandStreamReceiverHwTest,
using MI_LOAD_REGISTER_IMM = typename FamilyType::MI_LOAD_REGISTER_IMM;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
auto heaplessStateInit = commandStreamReceiver.heaplessStateInitialized;
bool preemptionModeProgramming = NEO::PreemptionHelper::getRequiredCmdStreamSize<FamilyType>(pDevice->getPreemptionMode(), commandStreamReceiver.getPreemptionMode()) > 0;
auto preemptionDetails = getPreemptionTestHwDetails<FamilyType>();
EXPECT_EQ(NEO::PreemptionMode::Initial, commandStreamReceiver.getPreemptionMode());
commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
if (!heaplessStateInit) {
EXPECT_EQ(NEO::PreemptionMode::Initial, commandStreamReceiver.getPreemptionMode());
}
heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
EXPECT_EQ(pDevice->getPreemptionMode(), commandStreamReceiver.getPreemptionMode());
HardwareParse hwParserCsr;
@@ -4592,10 +4682,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
EXPECT_EQ(preemptionModeProgramming, foundPreemptionMode);
size_t usedSize = commandStreamReceiver.commandStream.getUsed();
commandStreamReceiver.flushImmediateTask(commandStream,
commandStream.getUsed(),
immediateFlushTaskFlags,
*pDevice);
heaplessStateInit ? commandStreamReceiver.flushImmediateTaskStateless(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice)
: commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
hwParserCsr.tearDown();
hwParserCsr.parseCommands<FamilyType>(commandStreamReceiver.commandStream, usedSize);
@@ -4616,6 +4705,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
IsAtLeastXeHpCore) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.storeMakeResidentAllocations = true;
if (commandStreamReceiver.getPreemptionAllocation() == nullptr) {
auto createdPreemptionAllocation = commandStreamReceiver.createPreemptionAllocation();
@@ -4661,6 +4753,10 @@ HWTEST2_F(CommandStreamReceiverHwTest,
pDevice->getExecutionEnvironment()->rootDeviceEnvironments[0]->initDebuggerL0(pDevice);
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.storeMakeResidentAllocations = true;
EXPECT_FALSE(commandStreamReceiver.getSipSentFlag());
@@ -4717,6 +4813,9 @@ HWTEST2_F(CommandStreamReceiverHwTest,
auto &compilerProductHelper = pDevice->getCompilerProductHelper();
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.storeMakeResidentAllocations = true;
EXPECT_FALSE(commandStreamReceiver.getSipSentFlag());
@@ -5004,6 +5103,9 @@ HWTEST_F(CommandStreamReceiverHwTest, givenRequiresInstructionCacheFlushWhenFlus
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
commandStreamReceiver.registerInstructionCacheFlush();
this->requiredStreamProperties.stateComputeMode.setPropertiesAll(false, GrfConfig::defaultGrfNumber, ThreadArbitrationPolicy::AgeBased, NEO::PreemptionMode::ThreadGroup);

3
shared/test/unit_test/device/neo_device_tests.cpp
View File

@@ -1720,6 +1720,9 @@ TEST_F(DeviceTests, givenDebuggerRequestedByUserAndNotAvailableWhenDeviceIsIniti
TEST_F(DeviceTests, givenDebuggerRequestedByUserWhenDeviceWithSubDevicesCreatedThenInitializeDebuggerOncePerRootDevice) {
extern size_t createDebuggerCallCount;
DebugManagerStateRestore dbgRestorer;
debugManager.flags.DeferStateInitSubmissionToFirstRegularUsage.set(1);
createDebuggerCallCount = 0;
auto executionEnvironment = MockDevice::prepareExecutionEnvironment(defaultHwInfo.get(), 0u);
executionEnvironment->setDebuggingMode(DebuggingMode::online);

29
shared/test/unit_test/direct_submission/direct_submission_tests_2.cpp
View File

@@ -2192,7 +2192,12 @@ HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenBcsRelaxedOrderingEnabledWh
ultCsr->programEndingCmd(commandStream, &endingPtr, true, true);
auto lrrCmd = reinterpret_cast<MI_LOAD_REGISTER_REG *>(commandStream.getCpuBase());
HardwareParse hwParser;
hwParser.parseCommands<FamilyType>(commandStream, 0);
hwParser.findHardwareCommands<FamilyType>();
auto lrrCmdIt = find<MI_LOAD_REGISTER_REG *>(hwParser.cmdList.begin(), hwParser.cmdList.end());
auto lrrCmd = genCmdCast<MI_LOAD_REGISTER_REG *>(*lrrCmdIt);
EXPECT_EQ(lrrCmd->getSourceRegisterAddress(), RegisterOffsets::csGprR3);
EXPECT_EQ(lrrCmd->getDestinationRegisterAddress(), RegisterOffsets::csGprR0);
@@ -2220,12 +2225,16 @@ HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenBcsRelaxedOrderingDisabledW
ultCsr->blitterDirectSubmission.reset(directSubmission);
auto &commandStream = ultCsr->getCS(0x100);
auto endingPtr = commandStream.getSpace(0);
ultCsr->programEndingCmd(commandStream, &endingPtr, true, false);
auto bbStartCmd = genCmdCast<MI_BATCH_BUFFER_START *>(commandStream.getCpuBase());
HardwareParse hwParser;
hwParser.parseCommands<FamilyType>(commandStream, 0);
hwParser.findHardwareCommands<FamilyType>();
auto bbStartCmdIt = find<MI_BATCH_BUFFER_START *>(hwParser.cmdList.begin(), hwParser.cmdList.end());
auto bbStartCmd = genCmdCast<MI_BATCH_BUFFER_START *>(*bbStartCmdIt);
ASSERT_NE(nullptr, bbStartCmd);
EXPECT_EQ(0u, bbStartCmd->getIndirectAddressEnable());
}
@@ -2248,7 +2257,12 @@ HWTEST2_F(DirectSubmissionRelaxedOrderingTests, whenProgrammingEndingCmdsThenSet
ultCsr->programEndingCmd(commandStream, &endingPtr, true, true);
auto lrrCmd = reinterpret_cast<MI_LOAD_REGISTER_REG *>(commandStream.getCpuBase());
HardwareParse hwParser;
hwParser.parseCommands<FamilyType>(commandStream, 0);
hwParser.findHardwareCommands<FamilyType>();
auto lrrCmdIt = find<MI_LOAD_REGISTER_REG *>(hwParser.cmdList.begin(), hwParser.cmdList.end());
auto lrrCmd = genCmdCast<MI_LOAD_REGISTER_REG *>(*lrrCmdIt);
EXPECT_EQ(lrrCmd->getSourceRegisterAddress(), RegisterOffsets::csGprR3);
EXPECT_EQ(lrrCmd->getDestinationRegisterAddress(), RegisterOffsets::csGprR0);
@@ -2277,7 +2291,12 @@ HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenBbWithoutRelaxedOrderingDep
ultCsr->programEndingCmd(commandStream, &endingPtr, true, false);
auto bbStartCmd = genCmdCast<MI_BATCH_BUFFER_START *>(commandStream.getCpuBase());
HardwareParse hwParser;
hwParser.parseCommands<FamilyType>(commandStream, 0);
hwParser.findHardwareCommands<FamilyType>();
auto bbStartCmdIt = find<MI_BATCH_BUFFER_START *>(hwParser.cmdList.begin(), hwParser.cmdList.end());
auto bbStartCmd = genCmdCast<MI_BATCH_BUFFER_START *>(*bbStartCmdIt);
ASSERT_NE(nullptr, bbStartCmd);
EXPECT_EQ(0u, bbStartCmd->getIndirectAddressEnable());
}

7
shared/test/unit_test/encoders/test_encode_dispatch_kernel_pvc_and_later.cpp
View File

@@ -6,6 +6,7 @@
*/
#include "shared/source/command_stream/stream_properties.h"
#include "shared/source/helpers/compiler_product_helper.h"
#include "shared/source/helpers/in_order_cmd_helpers.h"
#include "shared/source/kernel/grf_config.h"
#include "shared/source/os_interface/product_helper.h"
@@ -86,6 +87,12 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTestPvcAndLater, givenCommandCon
MockExecutionEnvironment executionEnvironment{};
auto &rootDeviceEnvironment = *executionEnvironment.rootDeviceEnvironments[0];
auto &productHelper = rootDeviceEnvironment.getHelper<ProductHelper>();
auto &compilerProductHelper = rootDeviceEnvironment.getHelper<CompilerProductHelper>();
auto heaplessEnabled = compilerProductHelper.isHeaplessModeEnabled();
if (heaplessEnabled) {
GTEST_SKIP();
}
StreamProperties streamProperties{};
streamProperties.initSupport(rootDeviceEnvironment);
streamProperties.stateComputeMode.setPropertiesAll(false, GrfConfig::largeGrfNumber, 0u, PreemptionMode::Disabled);

2
shared/test/unit_test/execution_environment/execution_environment_tests.cpp
View File

@@ -384,6 +384,8 @@ TEST(ExecutionEnvironment, givenMultipleRootDevicesWhenTheyAreCreatedThenReuseMe
executionEnvironment.rootDeviceEnvironments[i]->setHwInfoAndInitHelpers(defaultHwInfo.get());
executionEnvironment.rootDeviceEnvironments[i]->initGmm();
}
executionEnvironment.calculateMaxOsContextCount();
std::unique_ptr<MockDevice> device(Device::create<MockDevice>(&executionEnvironment, 0u));
auto &commandStreamReceiver = device->getGpgpuCommandStreamReceiver();
auto memoryManager = device->getMemoryManager();

7
shared/test/unit_test/memory_manager/internal_allocation_storage_tests.cpp
View File

@@ -144,7 +144,12 @@ TEST_F(InternalAllocationStorageTest, whenCompletedAllocationIsStoredAsReusableA
memoryManager->freeGraphicsMemory(allocation);
}
TEST_F(InternalAllocationStorageTest, whenNotUsedAllocationIsStoredAsReusableAndThenCanBeObtained) {
HWTEST_F(InternalAllocationStorageTest, whenNotUsedAllocationIsStoredAsReusableAndThenCanBeObtained) {
auto ultCsr = reinterpret_cast<UltCommandStreamReceiver<FamilyType> *>(csr);
if (ultCsr->heaplessStateInitialized) {
ultCsr->taskCount.store(0);
}
auto allocation = memoryManager->allocateGraphicsMemoryWithProperties(AllocationProperties{0, MemoryConstants::pageSize, AllocationType::buffer, mockDeviceBitfield});
EXPECT_NE(nullptr, allocation);