Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Revert "Use GPU instead of CPU address in programming commands for HwTim(...)" 

This reverts commit 6202b2222b.
"Use GPU instead of CPU address in programming commands for HwTimeStamps"

Change-Id: I085382d95538ae41068a21c628d606039bf9cdf0
This commit is contained in:
Pawel Wilma 2018-12-20 16:32:47 +01:00 committed by sys_ocldev
parent 1e011f9a08
commit cc1f4bed60
14 changed files with 38 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
runtime/command_queue/enqueue_common.h
View File

@ -146,7 +146,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
auto devQueue = this->getContext().getDefaultDeviceQueue();
DeviceQueueHw<GfxFamily> *devQueueHw = castToObject<DeviceQueueHw<GfxFamily>>(devQueue);
TagNode<HwTimeStamps> *hwTimeStamps = nullptr;
HwTimeStamps *hwTimeStamps = nullptr;
auto commandStreamRecieverOwnership = getCommandStreamReceiver().obtainUniqueOwnership();
@ -230,7 +230,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
}
if (this->isProfilingEnabled()) {
// Get allocation for timestamps
hwTimeStamps = eventBuilder.getEvent()->getHwTimeStampNode();
hwTimeStamps = eventBuilder.getEvent()->getHwTimeStampNode()->tag;
if (this->isPerfCountersEnabled()) {
hwPerfCounter = eventBuilder.getEvent()->getHwPerfCounterNode()->tag;
// PERF COUNTER: copy current configuration from queue to event

4
runtime/command_queue/gpgpu_walker.h
View File

@ -139,11 +139,11 @@ class GpgpuWalkerHelper {
const iOpenCL::SPatchThreadPayload &threadPayload);
static void dispatchProfilingCommandsStart(
TagNode<HwTimeStamps> &hwTimeStamps,
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream);
static void dispatchProfilingCommandsEnd(
TagNode<HwTimeStamps> &hwTimeStamps,
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream);
static void dispatchPerfCountersNoopidRegisterCommands(

10
runtime/command_queue/gpgpu_walker.inl
View File

@ -101,17 +101,17 @@ void GpgpuWalkerHelper<GfxFamily>::addAluReadModifyWriteRegister(
template <typename GfxFamily>
void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
TagNode<HwTimeStamps> &hwTimeStamps,
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream) {
using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
// PIPE_CONTROL for global timestamp
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->GlobalStartTS, hwTimeStamps.tag);
uint64_t TimeStampAddress = reinterpret_cast<uint64_t>(&(hwTimeStamps.GlobalStartTS));
PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(commandStream, PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP, TimeStampAddress, 0llu);
//MI_STORE_REGISTER_MEM for context local timestamp
TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextStartTS, hwTimeStamps.tag);
TimeStampAddress = reinterpret_cast<uint64_t>(&(hwTimeStamps.ContextStartTS));
//low part
auto pMICmdLow = (MI_STORE_REGISTER_MEM *)commandStream->getSpace(sizeof(MI_STORE_REGISTER_MEM));
@ -122,7 +122,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
template <typename GfxFamily>
void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsEnd(
TagNode<HwTimeStamps> &hwTimeStamps,
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream) {
using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
@ -133,7 +133,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsEnd(
pPipeControlCmd->setCommandStreamerStallEnable(true);
//MI_STORE_REGISTER_MEM for context local timestamp
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextEndTS, hwTimeStamps.tag);
uint64_t TimeStampAddress = reinterpret_cast<uint64_t>(&(hwTimeStamps.ContextEndTS));
//low part
auto pMICmdLow = (MI_STORE_REGISTER_MEM *)commandStream->getSpace(sizeof(MI_STORE_REGISTER_MEM));

6
runtime/command_queue/hardware_interface.h
View File

@ -40,7 +40,7 @@ class HardwareInterface {
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
KernelOperation **blockedCommandsData,
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
TimestampPacketContainer *previousTimestampPacketNodes,
TimestampPacketContainer *currentTimestampPacketNodes,
@ -69,13 +69,13 @@ class HardwareInterface {
static void dispatchProfilingPerfStartCommands(
const DispatchInfo &dispatchInfo,
const MultiDispatchInfo &multiDispatchInfo,
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
LinearStream *commandStream,
CommandQueue &commandQueue);
static void dispatchProfilingPerfEndCommands(
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
LinearStream *commandStream,
CommandQueue &commandQueue);

2
runtime/command_queue/hardware_interface.inl
View File

@ -19,7 +19,7 @@ void HardwareInterface<GfxFamily>::dispatchWalker(
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
KernelOperation **blockedCommandsData,
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
TimestampPacketContainer *previousTimestampPacketNodes,
TimestampPacketContainer *currentTimestampPacketNodes,

4
runtime/command_queue/hardware_interface_base.inl
View File

@ -59,7 +59,7 @@ template <typename GfxFamily>
inline void HardwareInterface<GfxFamily>::dispatchProfilingPerfStartCommands(
const DispatchInfo &dispatchInfo,
const MultiDispatchInfo &multiDispatchInfo,
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
LinearStream *commandStream,
CommandQueue &commandQueue) {
@ -77,7 +77,7 @@ inline void HardwareInterface<GfxFamily>::dispatchProfilingPerfStartCommands(
template <typename GfxFamily>
inline void HardwareInterface<GfxFamily>::dispatchProfilingPerfEndCommands(
TagNode<HwTimeStamps> *hwTimeStamps,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
LinearStream *commandStream,
CommandQueue &commandQueue) {

4
runtime/device_queue/device_queue.cpp
View File

@ -143,7 +143,7 @@ void DeviceQueue::initDeviceQueue() {
igilEventPool->m_size = caps.maxOnDeviceEvents;
}
void DeviceQueue::setupExecutionModelDispatch(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentCount, uint32_t taskCount, TagNode<HwTimeStamps> *hwTimeStamp) {
void DeviceQueue::setupExecutionModelDispatch(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentCount, uint32_t taskCount, HwTimeStamps *hwTimeStamp) {
setupIndirectState(surfaceStateHeap, dynamicStateHeap, parentKernel, parentCount);
addExecutionModelCleanUpSection(parentKernel, hwTimeStamp, taskCount);
}
@ -152,7 +152,7 @@ void DeviceQueue::setupIndirectState(IndirectHeap &surfaceStateHeap, IndirectHea
return;
}
void DeviceQueue::addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) {
void DeviceQueue::addExecutionModelCleanUpSection(Kernel *parentKernel, HwTimeStamps *hwTimeStamp, uint32_t taskCount) {
return;
}

6
runtime/device_queue/device_queue.h
View File

@ -22,8 +22,6 @@ class Event;
struct MultiDispatchInfo;
class SchedulerKernel;
struct HwTimeStamps;
template <class T>
struct TagNode;
template <>
struct OpenCLObjectMapper<_device_queue> {
@ -68,10 +66,10 @@ class DeviceQueue : public BaseObject<_device_queue> {
size_t paramValueSize, void *paramValue,
size_t *paramValueSizeRet);
void setupExecutionModelDispatch(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentCount, uint32_t taskCount, TagNode<HwTimeStamps> *hwTimeStamp);
void setupExecutionModelDispatch(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentCount, uint32_t taskCount, HwTimeStamps *hwTimeStamp);
virtual void setupIndirectState(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentIDCount);
virtual void addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount);
virtual void addExecutionModelCleanUpSection(Kernel *parentKernel, HwTimeStamps *hwTimeStamp, uint32_t taskCount);
MOCKABLE_VIRTUAL bool isEMCriticalSectionFree() {
auto igilCmdQueue = reinterpret_cast<IGIL_CommandQueue *>(queueBuffer->getUnderlyingBuffer());

2
runtime/device_queue/device_queue_hw.h
View File

@ -55,7 +55,7 @@ class DeviceQueueHw : public DeviceQueue {
void setupIndirectState(IndirectHeap &surfaceStateHeap, IndirectHeap &dynamicStateHeap, Kernel *parentKernel, uint32_t parentIDCount) override;
void addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) override;
void addExecutionModelCleanUpSection(Kernel *parentKernel, HwTimeStamps *hwTimeStamp, uint32_t taskCount) override;
void resetDeviceQueue() override;
void dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override;

5
runtime/device_queue/device_queue_hw.inl
View File

@ -12,7 +12,6 @@
#include "runtime/helpers/preamble.h"
#include "runtime/helpers/string.h"
#include "runtime/memory_manager/memory_manager.h"
#include "runtime/utilities/tag_allocator.h"
namespace OCLRT {
template <typename GfxFamily>
@ -202,7 +201,7 @@ void DeviceQueueHw<GfxFamily>::buildSlbDummyCommands() {
}
template <typename GfxFamily>
void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) {
void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKernel, HwTimeStamps *hwTimeStamp, uint32_t taskCount) {
// CleanUp Section
auto offset = slbCS.getUsed();
auto alignmentSize = alignUp(offset, MemoryConstants::pageSize) - offset;
@ -216,7 +215,7 @@ void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKer
using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
if (hwTimeStamp != nullptr) {
uint64_t TimeStampAddress = hwTimeStamp->getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamp->tag->ContextCompleteTS, hwTimeStamp->tag);
uint64_t TimeStampAddress = (uint64_t)((uintptr_t) & (hwTimeStamp->ContextCompleteTS));
igilQueue->m_controls.m_EventTimestampAddress = TimeStampAddress;
addProfilingEndCmds(TimeStampAddress);

2
runtime/event/event.cpp
View File

@ -451,7 +451,7 @@ void Event::submitCommand(bool abortTasks) {
if ((this->isProfilingEnabled()) && (this->cmdQueue != nullptr)) {
if (timeStampNode) {
this->cmdQueue->getCommandStreamReceiver().makeResident(*timeStampNode->getGraphicsAllocation());
cmdToProcess->timestamp = timeStampNode;
cmdToProcess->timestamp = timeStampNode->tag;
}
if (profilingCpuPath) {
setSubmitTimeStamp();

4
runtime/helpers/task_information.h
View File

@ -26,8 +26,6 @@ class Surface;
class PrintfHandler;
struct HwTimeStamps;
class TimestampPacketContainer;
template <class T>
struct TagNode;
enum MapOperationType {
MAP,
@ -44,7 +42,7 @@ class Command : public IFNode<Command> {
virtual LinearStream *getCommandStream() {
return nullptr;
}
TagNode<HwTimeStamps> *timestamp = nullptr;
HwTimeStamps *timestamp = nullptr;
CompletionStamp completionStamp = {};
};

18
unit_tests/device_queue/device_queue_hw_tests.cpp
View File

@ -7,7 +7,6 @@
#include "hw_cmds.h"
#include "runtime/helpers/options.h"
#include "runtime/utilities/tag_allocator.h"
#include "unit_tests/fixtures/device_host_queue_fixture.h"
#include "unit_tests/fixtures/execution_model_fixture.h"
#include "unit_tests/helpers/hw_parse.h"
@ -353,19 +352,18 @@ HWCMDTEST_F(IGFX_GEN8_CORE, DeviceQueueSlb, AddEMCleanupSectionWithProfiling) {
MockParentKernel *mockParentKernel = MockParentKernel::create(*pContext);
uint32_t taskCount = 7;
auto hwTimeStamp = pCommandQueue->getCommandStreamReceiver().getEventTsAllocator()->getTag();
HwTimeStamps hwTimeStamp;
mockDeviceQueueHw->buildSlbDummyCommands();
mockDeviceQueueHw->addExecutionModelCleanUpSection(mockParentKernel, hwTimeStamp, taskCount);
mockDeviceQueueHw->addExecutionModelCleanUpSection(mockParentKernel, &hwTimeStamp, taskCount);
uint32_t eventTimestampAddrLow = static_cast<uint32_t>(igilCmdQueue->m_controls.m_EventTimestampAddress & 0xFFFFFFFF);
uint32_t eventTimestampAddrHigh = static_cast<uint32_t>((igilCmdQueue->m_controls.m_EventTimestampAddress & 0xFFFFFFFF00000000) >> 32);
uint32_t eventTimestampLow = (uint32_t)(igilCmdQueue->m_controls.m_EventTimestampAddress & 0xFFFFFFFF);
uint32_t eventTimestampHigh = (uint32_t)((igilCmdQueue->m_controls.m_EventTimestampAddress & 0xFFFFFFFF00000000) >> 32);
uint64_t contextCompleteAddr = hwTimeStamp->getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamp->tag->ContextCompleteTS, hwTimeStamp->tag);
uint32_t contextCompleteAddrLow = static_cast<uint32_t>(contextCompleteAddr & 0xFFFFFFFF);
uint32_t contextCompleteAddrHigh = static_cast<uint32_t>((contextCompleteAddr & 0xFFFFFFFF00000000) >> 32);
uint32_t contextCompleteLow = (uint32_t)((uint64_t)((uintptr_t)(&hwTimeStamp.ContextCompleteTS)) & 0xFFFFFFFF);
uint32_t contextCompleteHigh = (uint32_t)(((uint64_t)((uintptr_t)(&hwTimeStamp.ContextCompleteTS)) & 0xFFFFFFFF00000000) >> 32);
EXPECT_EQ(contextCompleteAddrLow, eventTimestampAddrLow);
EXPECT_EQ(contextCompleteAddrHigh, eventTimestampAddrHigh);
EXPECT_EQ(contextCompleteLow, eventTimestampLow);
EXPECT_EQ(contextCompleteHigh, eventTimestampHigh);
HardwareParse hwParser;
auto *slbCS = mockDeviceQueueHw->getSlbCS();

12
unit_tests/execution_model/submit_blocked_parent_kernel_tests.cpp
View File

@ -8,7 +8,6 @@
#include "runtime/command_queue/gpgpu_walker.h"
#include "runtime/command_queue/hardware_interface.h"
#include "runtime/event/hw_timestamps.h"
#include "runtime/utilities/tag_allocator.h"
#include "runtime/helpers/kernel_commands.h"
#include "runtime/helpers/task_information.h"
#include "unit_tests/mocks/mock_command_queue.h"
@ -57,9 +56,9 @@ class MockDeviceQueueHwWithCriticalSectionRelease : public DeviceQueueHw<GfxFami
indirectStateSetup = true;
return BaseClass::setupIndirectState(surfaceStateHeap, dynamicStateHeap, parentKernel, parentIDCount);
}
void addExecutionModelCleanUpSection(Kernel *parentKernel, TagNode<HwTimeStamps> *hwTimeStamp, uint32_t taskCount) override {
void addExecutionModelCleanUpSection(Kernel *parentKernel, HwTimeStamps *hwTimeStamp, uint32_t taskCount) override {
cleanupSectionAdded = true;
timestampAddedInCleanupSection = hwTimeStamp ? hwTimeStamp->tag : nullptr;
timestampAddedInCleanupSection = hwTimeStamp;
return BaseClass::addExecutionModelCleanUpSection(parentKernel, hwTimeStamp, taskCount);
}
void dispatchScheduler(CommandQueue &cmdQ, SchedulerKernel &scheduler, PreemptionMode preemptionMode, IndirectHeap *ssh, IndirectHeap *dsh) override {
@ -250,12 +249,13 @@ HWTEST_F(ParentKernelCommandQueueFixture, givenBlockedParentKernelWithProfilingW
std::vector<Surface *> surfaces;
auto *cmdComputeKernel = new CommandComputeKernel(*pCmdQ, std::unique_ptr<KernelOperation>(blockedCommandData), surfaces, false, false, false, nullptr, preemptionMode, parentKernel, 1);
auto timestamp = pCmdQ->getCommandStreamReceiver().getEventTsAllocator()->getTag();
cmdComputeKernel->timestamp = timestamp;
HwTimeStamps timestamp;
cmdComputeKernel->timestamp = &timestamp;
cmdComputeKernel->submit(0, false);
EXPECT_TRUE(mockDevQueue.cleanupSectionAdded);
EXPECT_EQ(mockDevQueue.timestampAddedInCleanupSection, timestamp->tag);
EXPECT_EQ(mockDevQueue.timestampAddedInCleanupSection, &timestamp);
delete cmdComputeKernel;
delete parentKernel;