intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-08 22:12:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Flush task at device init

Browse Source 
Resolves: NEO-7642

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2023-01-27 12:37:09 +00:00
committed by Compute-Runtime-Automation
parent d2628babbc
commit 2abbd82195
16 changed files with 81 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
level_zero/core/source/driver/driver.cpp
View File

@@ -53,6 +53,8 @@ void DriverImp::initialize(ze_result_t *result) {
}
}
executionEnvironment->setMetricsEnabled(envVariables.metrics);
executionEnvironment->incRefInternal();
auto neoDevices = NEO::DeviceFactory::createDevices(*executionEnvironment);
executionEnvironment->decRefInternal();

11
opencl/test/unit_test/command_queue/blit_enqueue_1_tests.cpp
View File

@@ -238,6 +238,17 @@ HWTEST_TEMPLATED_F(BlitAuxTranslationTests, whenFlushTagUpdateThenSetStallingCmd
EXPECT_TRUE(ultCsr->latestFlushedBatchBuffer.hasStallingCmds);
}
HWTEST_TEMPLATED_F(BlitAuxTranslationTests, whenInitializeDeviceWithFirstSubmissionThenMiFlushDwIsFlushed) {
using MI_FLUSH_DW = typename FamilyType::MI_FLUSH_DW;
EXPECT_EQ(SubmissionStatus::SUCCESS, bcsCsr->initializeDeviceWithFirstSubmission());
auto cmdListBcs = getCmdList<FamilyType>(bcsCsr->getCS(0), 0);
auto cmdFound = expectCommand<MI_FLUSH_DW>(cmdListBcs.begin(), cmdListBcs.end());
EXPECT_NE(cmdFound, cmdListBcs.end());
}
HWTEST_TEMPLATED_F(BlitAuxTranslationTests, givenBlitTranslationWhenConstructingCommandBufferThenSynchronizeBcsOutput) {
using XY_COPY_BLT = typename FamilyType::XY_COPY_BLT;
using MI_FLUSH_DW = typename FamilyType::MI_FLUSH_DW;

8
opencl/test/unit_test/linux/main_linux_dll.cpp
View File

@@ -884,6 +884,14 @@ TEST(DeviceTest, whenCheckBlitSplitEnabledThenReturnsTrue) {
EXPECT_TRUE(Device::isBlitSplitEnabled());
}
TEST(DeviceTest, givenCsrHwWhenCheckIsInitDeviceWithFirstSubmissionEnabledThenReturnsTrue) {
EXPECT_TRUE(Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType::CSR_HW));
}
TEST(DeviceTest, givenCsrNonHwWhenCheckIsInitDeviceWithFirstSubmissionEnabledThenReturnsTrue) {
EXPECT_FALSE(Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType::CSR_TBX));
}
TEST(PlatformsDestructor, whenGlobalPlatformsDestructorIsCalledThenGlobalPlatformsAreDestroyed) {
EXPECT_NE(nullptr, platformsImpl);
platformsDestructor();

13
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -1346,9 +1346,13 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushPipeControl() {
args.notifyEnable = isUsedNotifyEnableForPostSync();
args.workloadPartitionOffset = isMultiTileOperationEnabled();
auto &commandStream = getCS(MemorySynchronizationCommands<GfxFamily>::getSizeForBarrierWithPostSyncOperation(peekRootDeviceEnvironment(), args.tlbInvalidation));
auto dispatchSize = MemorySynchronizationCommands<GfxFamily>::getSizeForBarrierWithPostSyncOperation(peekRootDeviceEnvironment(), args.tlbInvalidation) + this->getCmdSizeForPrologue();
auto &commandStream = getCS(dispatchSize);
auto commandStreamStart = commandStream.getUsed();
this->programEnginePrologue(commandStream);
MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(commandStream,
PostSyncMode::ImmediateData,
getTagAllocation()->getGpuAddress(),
@@ -1357,6 +1361,7 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushPipeControl() {
args);
makeResident(*tagAllocation);
makeResident(*commandStream.getGraphicsAllocation());
auto submissionStatus = this->flushSmallTask(commandStream, commandStreamStart);
this->latestFlushedTaskCount = taskCount.load();
@@ -1594,11 +1599,7 @@ void CommandStreamReceiverHw<GfxFamily>::createKernelArgsBufferAllocation() {
template <typename GfxFamily>
SubmissionStatus CommandStreamReceiverHw<GfxFamily>::initializeDeviceWithFirstSubmission() {
auto lock = obtainUniqueOwnership();
auto &commandStream = getCS(EncodeBatchBufferStartOrEnd<GfxFamily>::getBatchBufferEndSize());
auto commandStreamStart = commandStream.getUsed();
return this->flushSmallTask(commandStream, commandStreamStart);
return flushTagUpdate();
}
template <typename GfxFamily>

18
shared/source/device/device.cpp
View File

@@ -383,13 +383,18 @@ bool Device::createEngine(uint32_t deviceCsrIndex, EngineTypeUsage engineTypeUsa
commandStreamReceiver->createKernelArgsBufferAllocation();
if (preemptionMode == PreemptionMode::MidThread && !commandStreamReceiver->createPreemptionAllocation()) {
return false;
}
if (isDefaultEngine) {
bool defaultEngineAlreadySet = (allEngines.size() > defaultEngineIndex) && (allEngines[defaultEngineIndex].getEngineType() == engineType);
if (!defaultEngineAlreadySet) {
defaultEngineIndex = deviceCsrIndex;
if (osContext->isDebuggableContext()) {
if (osContext->isDebuggableContext() ||
this->isInitDeviceWithFirstSubmissionSupported(commandStreamReceiver->getType())) {
if (SubmissionStatus::SUCCESS != commandStreamReceiver->initializeDeviceWithFirstSubmission()) {
return false;
}
@@ -401,15 +406,12 @@ bool Device::createEngine(uint32_t deviceCsrIndex, EngineTypeUsage engineTypeUsa
defaultBcsEngineIndex = deviceCsrIndex;
}
if (preemptionMode == PreemptionMode::MidThread && !commandStreamReceiver->createPreemptionAllocation()) {
return false;
}
EngineControl engine{commandStreamReceiver.get(), osContext};
allEngines.push_back(engine);
if (engineUsage == EngineUsage::Regular) {
addEngineToEngineGroup(engine);
}
commandStreamReceivers.push_back(std::move(commandStreamReceiver));
return true;
@@ -442,6 +444,12 @@ bool Device::isBcsSplitSupported() {
return bcsSplit;
}
bool Device::isInitDeviceWithFirstSubmissionSupported(CommandStreamReceiverType csrType) {
return !this->executionEnvironment->areMetricsEnabled() &&
getProductHelper().isInitDeviceWithFirstSubmissionRequired(getHardwareInfo()) &&
Device::isInitDeviceWithFirstSubmissionEnabled(csrType);
}
double Device::getPlatformHostTimerResolution() const {
if (getOSTime()) {
return getOSTime()->getHostTimerResolution();

3
shared/source/device/device.h
View File

@@ -10,6 +10,7 @@
#include "shared/source/helpers/engine_control.h"
#include "shared/source/helpers/engine_node_helper.h"
#include "shared/source/helpers/non_copyable_or_moveable.h"
#include "shared/source/helpers/options.h"
#include "shared/source/os_interface/hw_info_config.h"
#include "shared/source/os_interface/performance_counters.h"
#include "shared/source/utilities/reference_tracked_object.h"
@@ -108,7 +109,9 @@ class Device : public ReferenceTrackedObject<Device> {
RootDeviceEnvironment &getRootDeviceEnvironmentRef() const;
bool isFullRangeSvm() const;
static bool isBlitSplitEnabled();
static bool isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType);
bool isBcsSplitSupported();
bool isInitDeviceWithFirstSubmissionSupported(CommandStreamReceiverType csrType);
bool areSharedSystemAllocationsAllowed() const;
template <typename SpecializedDeviceT>
void setSpecializedDevice(SpecializedDeviceT *specializedDevice) {

4
shared/source/dll/device_dll.cpp
View File

@@ -13,4 +13,8 @@ bool Device::isBlitSplitEnabled() {
return true;
}
bool Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType) {
return csrType == CommandStreamReceiverType::CSR_HW;
}
} // namespace NEO

7
shared/source/execution_environment/execution_environment.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2022 Intel Corporation
* Copyright (C) 2018-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -38,6 +38,10 @@ class ExecutionEnvironment : public ReferenceTrackedObject<ExecutionEnvironment>
debuggingEnabled = true;
}
bool isDebuggingEnabled() { return debuggingEnabled; }
void setMetricsEnabled(bool value) {
this->metricsEnabled = value;
}
bool areMetricsEnabled() { return this->metricsEnabled; }
DirectSubmissionController *initializeDirectSubmissionController();
std::unique_ptr<MemoryManager> memoryManager;
@@ -51,6 +55,7 @@ class ExecutionEnvironment : public ReferenceTrackedObject<ExecutionEnvironment>
void adjustCcsCountImpl(RootDeviceEnvironment *rootDeviceEnvironment) const;
void configureNeoEnvironment();
bool debuggingEnabled = false;
bool metricsEnabled = false;
std::unordered_map<uint32_t, uint32_t> rootDeviceNumCcsMap;
};
} // namespace NEO

2
shared/source/helpers/completion_stamp.h
View File

@@ -7,6 +7,7 @@
#pragma once
#include "shared/source/command_stream/submission_status.h"
#include "shared/source/command_stream/task_count_helper.h"
#include <cstdint>
@@ -14,7 +15,6 @@
namespace NEO {
using FlushStamp = uint64_t;
enum class SubmissionStatus : uint32_t;
struct CompletionStamp {
static TaskCountType getTaskCountFromSubmissionStatusError(SubmissionStatus submissionStatus);

2
shared/source/os_interface/hw_info_config.h
View File

@@ -136,6 +136,7 @@ class ProductHelper {
virtual bool isBcsReportWaRequired(const HardwareInfo &hwInfo) const = 0;
virtual bool isBlitSplitEnqueueWARequired(const HardwareInfo &hwInfo) const = 0;
virtual bool isBlitCopyRequiredForLocalMemory(const RootDeviceEnvironment &rootDeviceEnvironment, const GraphicsAllocation &allocation) const = 0;
virtual bool isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const = 0;
virtual bool isImplicitScalingSupported(const HardwareInfo &hwInfo) const = 0;
virtual bool isCpuCopyNecessary(const void *ptr, MemoryManager *memoryManager) const = 0;
virtual bool isAdjustWalkOrderAvailable(const HardwareInfo &hwInfo) const = 0;
@@ -280,6 +281,7 @@ class ProductHelperHw : public ProductHelper {
bool isTimestampWaitSupportedForEvents() const override;
bool isTilePlacementResourceWaRequired(const HardwareInfo &hwInfo) const override;
bool isBlitSplitEnqueueWARequired(const HardwareInfo &hwInfo) const override;
bool isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const override;
bool allowMemoryPrefetch(const HardwareInfo &hwInfo) const override;
bool isBcsReportWaRequired(const HardwareInfo &hwInfo) const override;
bool isBlitCopyRequiredForLocalMemory(const RootDeviceEnvironment &rootDeviceEnvironment, const GraphicsAllocation &allocation) const override;

5
shared/source/os_interface/hw_info_config.inl
View File

@@ -479,6 +479,11 @@ bool ProductHelperHw<gfxProduct>::isBlitCopyRequiredForLocalMemory(const RootDev
!allocation.isAllocationLockable());
}
template <PRODUCT_FAMILY gfxProduct>
bool ProductHelperHw<gfxProduct>::isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const {
return false;
}
template <PRODUCT_FAMILY gfxProduct>
bool ProductHelperHw<gfxProduct>::isImplicitScalingSupported(const HardwareInfo &hwInfo) const {
return false;

5
shared/source/xe_hpc_core/pvc/os_agnostic_hw_info_config_pvc.inl
View File

@@ -176,6 +176,11 @@ bool ProductHelperHw<gfxProduct>::isBlitSplitEnqueueWARequired(const HardwareInf
return true;
}
template <>
bool ProductHelperHw<gfxProduct>::isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const {
return true;
}
template <>
bool ProductHelperHw<gfxProduct>::isImplicitScalingSupported(const HardwareInfo &hwInfo) const {
return getSteppingFromHwRevId(hwInfo) >= REVISION_B;

2
shared/test/common/base_ult_config_listener.cpp
View File

@@ -36,7 +36,7 @@ void NEO::BaseUltConfigListener::OnTestEnd(const ::testing::TestInfo &) {
// Ensure that global state is restored
UltHwConfig expectedState{};
static_assert(sizeof(UltHwConfig) == 13 * sizeof(bool), ""); // Ensure that there is no internal padding
static_assert(sizeof(UltHwConfig) == 14 * sizeof(bool), ""); // Ensure that there is no internal padding
EXPECT_EQ(0, memcmp(&expectedState, &ultHwConfig, sizeof(UltHwConfig)));
EXPECT_EQ(0, memcmp(&referencedHwInfo.platform, &defaultHwInfo->platform, sizeof(PLATFORM)));

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

@@ -14,6 +14,7 @@ struct UltHwConfig {
bool forceOsAgnosticMemoryManager = true;
bool useWaitForTimestamps = false;
bool useBlitSplit = false;
bool useFirstSubmissionInitDevice = false;
bool csrFailInitDirectSubmission = false;
bool csrBaseCallDirectSubmissionAvailable = false;

5
shared/test/common/mocks/device_ult.cpp
View File

@@ -7,11 +7,14 @@
#include "shared/source/device/device.h"
#include "shared/test/common/helpers/ult_hw_config.h"
namespace NEO {
bool Device::isBlitSplitEnabled() {
return ultHwConfig.useBlitSplit;
}
bool Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType) {
return ultHwConfig.useFirstSubmissionInitDevice;
}
} // namespace NEO

10
shared/test/unit_test/device/neo_device_tests.cpp
View File

@@ -673,6 +673,9 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenZexNumberOfCssEnvVariableSetAmbig
}
HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenDebuggableOsContextWhenDeviceCreatesEnginesThenDeviceIsInitializedWithFirstSubmission) {
VariableBackup<UltHwConfig> backup(&ultHwConfig);
ultHwConfig.useFirstSubmissionInitDevice = true;
auto hwInfo = *defaultHwInfo;
hardwareInfoSetup[hwInfo.platform.eProductFamily](&hwInfo, true, 0);
@@ -687,7 +690,10 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenDebuggableOsContextWhenDeviceCrea
EXPECT_EQ(1u, csr->peekLatestSentTaskCount());
}
HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenNonDebuggableOsContextWhenDeviceCreatesEnginesThenDeviceIsNotInitializedWithFirstSubmission) {
HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, whenDeviceCreatesEnginesThenDeviceIsInitializedWithFirstSubmission) {
VariableBackup<UltHwConfig> backup(&ultHwConfig);
ultHwConfig.useFirstSubmissionInitDevice = true;
auto hwInfo = *defaultHwInfo;
hardwareInfoSetup[hwInfo.platform.eProductFamily](&hwInfo, true, 0);
@@ -698,7 +704,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenNonDebuggableOsContextWhenDeviceC
auto device = deviceFactory.rootDevices[0];
auto csr = device->allEngines[device->defaultEngineIndex].commandStreamReceiver;
EXPECT_EQ(0u, csr->peekLatestSentTaskCount());
EXPECT_EQ(device->isInitDeviceWithFirstSubmissionSupported(csr->getType()), csr->peekLatestSentTaskCount());
}
TEST(FailDeviceTest, GivenFailedDeviceWhenCreatingDeviceThenNullIsReturned) {