compute-runtime/runtime/event/event.cpp

693 lines
24 KiB
C++
Raw Normal View History

/*
* Copyright (C) 2017-2018 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "public/cl_ext_private.h"
#include "runtime/command_queue/command_queue.h"
#include "runtime/command_stream/command_stream_receiver.h"
#include "runtime/memory_manager/internal_allocation_storage.h"
#include "runtime/context/context.h"
#include "runtime/device/device.h"
#include "runtime/event/event.h"
#include "runtime/event/event_tracker.h"
#include "runtime/helpers/aligned_memory.h"
#include "runtime/helpers/get_info.h"
#include "runtime/helpers/kernel_commands.h"
#include "runtime/helpers/timestamp_packet.h"
#include "runtime/api/cl_types.h"
#include "runtime/mem_obj/mem_obj.h"
#include "runtime/utilities/range.h"
#include "runtime/utilities/stackvec.h"
#include "runtime/utilities/tag_allocator.h"
#include "runtime/platform/platform.h"
#include "runtime/event/async_events_handler.h"
namespace OCLRT {
const cl_uint Event::eventNotReady = 0xFFFFFFF0;
Event::Event(
Context *ctx,
CommandQueue *cmdQueue,
cl_command_type cmdType,
uint32_t taskLevel,
uint32_t taskCount)
: taskLevel(taskLevel),
currentCmdQVirtualEvent(false),
cmdToSubmit(nullptr),
submittedCmd(nullptr),
ctx(ctx),
cmdQueue(cmdQueue),
cmdType(cmdType),
dataCalculated(false),
taskCount(taskCount) {
if (OCLRT::DebugManager.flags.EventsTrackerEnable.get()) {
EventsTracker::getEventsTracker().notifyCreation(this);
}
parentCount = 0;
executionStatus = CL_QUEUED;
flushStamp.reset(new FlushStampTracker(true));
DBG_LOG(EventsDebugEnable, "Event()", this);
// Event can live longer than command queue that created it,
// hence command queue refCount must be incremented
// non-null command queue is only passed when Base Event object is created
// any other Event types must increment refcount when setting command queue
if (cmdQueue != nullptr) {
cmdQueue->incRefInternal();
}
if ((this->ctx == nullptr) && (cmdQueue != nullptr)) {
this->ctx = &cmdQueue->getContext();
if (cmdQueue->getCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
timestampPacketContainer = std::make_unique<TimestampPacketContainer>();
}
}
if (this->ctx != nullptr) {
this->ctx->incRefInternal();
}
queueTimeStamp = {0, 0};
submitTimeStamp = {0, 0};
startTimeStamp = 0;
endTimeStamp = 0;
completeTimeStamp = 0;
profilingEnabled = !isUserEvent() &&
(cmdQueue ? cmdQueue->getCommandQueueProperties() & CL_QUEUE_PROFILING_ENABLE : false);
profilingCpuPath = ((cmdType == CL_COMMAND_MAP_BUFFER) || (cmdType == CL_COMMAND_MAP_IMAGE)) && profilingEnabled;
perfCountersEnabled = cmdQueue ? cmdQueue->isPerfCountersEnabled() : false;
}
Event::Event(
CommandQueue *cmdQueue,
cl_command_type cmdType,
uint32_t taskLevel,
uint32_t taskCount)
: Event(nullptr, cmdQueue, cmdType, taskLevel, taskCount) {
}
Event::~Event() {
if (OCLRT::DebugManager.flags.EventsTrackerEnable.get()) {
EventsTracker::getEventsTracker().notifyDestruction(this);
}
DBG_LOG(EventsDebugEnable, "~Event()", this);
//no commands should be registred
DEBUG_BREAK_IF(this->cmdToSubmit.load());
submitCommand(true);
int32_t lastStatus = executionStatus;
if (isStatusCompleted(&lastStatus) == false) {
transitionExecutionStatus(-1);
DEBUG_BREAK_IF(peekHasCallbacks() || peekHasChildEvents());
}
// Note from OCL spec:
// "All callbacks registered for an event object must be called.
// All enqueued callbacks shall be called before the event object is destroyed."
if (peekHasCallbacks()) {
executeCallbacks(lastStatus);
}
{
// clean-up submitted command if needed
std::unique_ptr<Command> submittedCommand(submittedCmd.exchange(nullptr));
}
if (cmdQueue != nullptr) {
if (timeStampNode != nullptr) {
timeStampNode->returnTag();
}
if (perfCounterNode != nullptr) {
perfCounterNode->returnTag();
}
cmdQueue->decRefInternal();
}
if (ctx != nullptr) {
ctx->decRefInternal();
}
if (perfConfigurationData) {
delete perfConfigurationData;
}
// in case event did not unblock child events before
unblockEventsBlockedByThis(executionStatus);
}
cl_int Event::getEventProfilingInfo(cl_profiling_info paramName,
size_t paramValueSize,
void *paramValue,
size_t *paramValueSizeRet) {
cl_int retVal;
const void *src = nullptr;
size_t srcSize = 0;
// CL_PROFILING_INFO_NOT_AVAILABLE if event refers to the clEnqueueSVMFree command
if (isUserEvent() != CL_FALSE || // or is a user event object.
!updateStatusAndCheckCompletion() || //if the execution status of the command identified by event is not CL_COMPLETE
!profilingEnabled) // the CL_QUEUE_PROFILING_ENABLE flag is not set for the command-queue,
{
return CL_PROFILING_INFO_NOT_AVAILABLE;
}
// if paramValue is NULL, it is ignored
switch (paramName) {
case CL_PROFILING_COMMAND_QUEUED:
src = &queueTimeStamp.CPUTimeinNS;
if (DebugManager.flags.ReturnRawGpuTimestamps.get()) {
src = &queueTimeStamp.GPUTimeStamp;
}
srcSize = sizeof(cl_ulong);
break;
case CL_PROFILING_COMMAND_SUBMIT:
src = &submitTimeStamp.CPUTimeinNS;
if (DebugManager.flags.ReturnRawGpuTimestamps.get()) {
src = &submitTimeStamp.GPUTimeStamp;
}
srcSize = sizeof(cl_ulong);
break;
case CL_PROFILING_COMMAND_START:
calcProfilingData();
src = &startTimeStamp;
srcSize = sizeof(cl_ulong);
break;
case CL_PROFILING_COMMAND_END:
calcProfilingData();
src = &endTimeStamp;
srcSize = sizeof(cl_ulong);
break;
case CL_PROFILING_COMMAND_COMPLETE:
calcProfilingData();
src = &completeTimeStamp;
srcSize = sizeof(cl_ulong);
break;
case CL_PROFILING_COMMAND_PERFCOUNTERS_INTEL:
if (!perfCountersEnabled) {
return CL_INVALID_VALUE;
}
if (!cmdQueue->getPerfCounters()->processEventReport(paramValueSize,
paramValue,
paramValueSizeRet,
getHwPerfCounterNode()->tag,
perfConfigurationData,
updateStatusAndCheckCompletion())) {
return CL_PROFILING_INFO_NOT_AVAILABLE;
}
return CL_SUCCESS;
default:
return CL_INVALID_VALUE;
}
retVal = ::getInfo(paramValue, paramValueSize, src, srcSize);
if (paramValueSizeRet) {
*paramValueSizeRet = srcSize;
}
return retVal;
} // namespace OCLRT
uint32_t Event::getCompletionStamp() const {
return this->taskCount;
}
void Event::updateCompletionStamp(uint32_t taskCount, uint32_t tasklevel, FlushStamp flushStamp) {
this->taskCount = taskCount;
this->taskLevel = tasklevel;
this->flushStamp->setStamp(flushStamp);
}
cl_ulong Event::getDelta(cl_ulong startTime,
cl_ulong endTime) {
cl_ulong Max = (1ULL << OCLRT_NUM_TIMESTAMP_BITS) - 1;
cl_ulong Delta = 0;
startTime &= Max;
endTime &= Max;
if (startTime > endTime) {
Delta = Max - startTime;
Delta += endTime;
} else {
Delta = endTime - startTime;
}
return Delta;
}
bool Event::calcProfilingData() {
uint64_t gpuDuration = 0;
uint64_t cpuDuration = 0;
uint64_t gpuCompleteDuration = 0;
uint64_t cpuCompleteDuration = 0;
int64_t c0 = 0;
if (!dataCalculated && timeStampNode && !profilingCpuPath) {
double frequency = cmdQueue->getDevice().getDeviceInfo().profilingTimerResolution;
/* calculation based on equation
CpuTime = GpuTime * scalar + const( == c0)
scalar = DeltaCpu( == dCpu) / DeltaGpu( == dGpu)
to determine the value of the const we can use one pair of values
const = CpuTimeQueue - GpuTimeQueue * scalar
*/
//If device enqueue has not updated complete timestamp, assign end timestamp
if (((HwTimeStamps *)timeStampNode->tag)->ContextCompleteTS == 0)
((HwTimeStamps *)timeStampNode->tag)->ContextCompleteTS = ((HwTimeStamps *)timeStampNode->tag)->ContextEndTS;
c0 = queueTimeStamp.CPUTimeinNS - static_cast<uint64_t>(queueTimeStamp.GPUTimeStamp * frequency);
gpuDuration = getDelta(
((HwTimeStamps *)timeStampNode->tag)->ContextStartTS,
((HwTimeStamps *)timeStampNode->tag)->ContextEndTS);
gpuCompleteDuration = getDelta(
((HwTimeStamps *)timeStampNode->tag)->ContextStartTS,
((HwTimeStamps *)timeStampNode->tag)->ContextCompleteTS);
cpuDuration = static_cast<uint64_t>(gpuDuration * frequency);
cpuCompleteDuration = static_cast<uint64_t>(gpuCompleteDuration * frequency);
startTimeStamp = static_cast<uint64_t>(((HwTimeStamps *)timeStampNode->tag)->GlobalStartTS * frequency) + c0;
endTimeStamp = startTimeStamp + cpuDuration;
completeTimeStamp = startTimeStamp + cpuCompleteDuration;
if (DebugManager.flags.ReturnRawGpuTimestamps.get()) {
startTimeStamp = ((HwTimeStamps *)timeStampNode->tag)->ContextStartTS;
endTimeStamp = ((HwTimeStamps *)timeStampNode->tag)->ContextEndTS;
completeTimeStamp = ((HwTimeStamps *)timeStampNode->tag)->ContextCompleteTS;
}
dataCalculated = true;
}
return dataCalculated;
}
inline bool Event::wait(bool blocking, bool useQuickKmdSleep) {
while (this->taskCount == Event::eventNotReady) {
if (blocking == false) {
return false;
}
}
cmdQueue->waitUntilComplete(taskCount.load(), flushStamp->peekStamp(), useQuickKmdSleep);
updateExecutionStatus();
DEBUG_BREAK_IF(this->taskLevel == Event::eventNotReady && this->executionStatus >= 0);
auto *allocationStorage = cmdQueue->getCommandStreamReceiver().getInternalAllocationStorage();
allocationStorage->cleanAllocationList(this->taskCount, TEMPORARY_ALLOCATION);
return true;
}
void Event::updateExecutionStatus() {
if (taskLevel == Event::eventNotReady) {
return;
}
int32_t statusSnapshot = executionStatus;
if (isStatusCompleted(&statusSnapshot)) {
executeCallbacks(statusSnapshot);
return;
}
if (peekIsBlocked()) {
transitionExecutionStatus(CL_QUEUED);
executeCallbacks(CL_QUEUED);
return;
}
if (statusSnapshot == CL_QUEUED) {
bool abortBlockedTasks = isStatusCompletedByTermination(&statusSnapshot);
submitCommand(abortBlockedTasks);
transitionExecutionStatus(CL_SUBMITTED);
executeCallbacks(CL_SUBMITTED);
unblockEventsBlockedByThis(CL_SUBMITTED);
// Note : Intentional fallthrough (no return) to check for CL_COMPLETE
}
if ((cmdQueue != nullptr) && (cmdQueue->isCompleted(getCompletionStamp()))) {
transitionExecutionStatus(CL_COMPLETE);
executeCallbacks(CL_COMPLETE);
unblockEventsBlockedByThis(CL_COMPLETE);
auto *allocationStorage = cmdQueue->getCommandStreamReceiver().getInternalAllocationStorage();
allocationStorage->cleanAllocationList(this->taskCount, TEMPORARY_ALLOCATION);
return;
}
transitionExecutionStatus(CL_SUBMITTED);
}
void Event::addChild(Event &childEvent) {
childEvent.parentCount++;
childEvent.incRefInternal();
childEventsToNotify.pushRefFrontOne(childEvent);
DBG_LOG(EventsDebugEnable, "addChild: Parent event:", this, "child:", &childEvent);
if (DebugManager.flags.TrackParentEvents.get()) {
childEvent.parentEvents.push_back(this);
}
if (executionStatus == CL_COMPLETE) {
unblockEventsBlockedByThis(CL_COMPLETE);
}
}
void Event::unblockEventsBlockedByThis(int32_t transitionStatus) {
int32_t status = transitionStatus;
(void)status;
DEBUG_BREAK_IF(!(isStatusCompleted(&status) || (peekIsSubmitted(&status))));
uint32_t taskLevelToPropagate = Event::eventNotReady;
if (isStatusCompletedByTermination(&transitionStatus) == false) {
//if we are event on top of the tree , obtain taskLevel from CSR
if (taskLevel == Event::eventNotReady) {
this->taskLevel = getTaskLevel();
taskLevelToPropagate = this->taskLevel;
} else {
taskLevelToPropagate = taskLevel + 1;
}
}
auto childEventRef = childEventsToNotify.detachNodes();
while (childEventRef != nullptr) {
auto childEvent = childEventRef->ref;
childEvent->unblockEventBy(*this, taskLevelToPropagate, transitionStatus);
if (childEvent->getCommandQueue() && childEvent->isCurrentCmdQVirtualEvent()) {
// Check virtual event state and delete it if possible.
childEvent->getCommandQueue()->isQueueBlocked();
}
childEvent->decRefInternal();
auto next = childEventRef->next;
delete childEventRef;
childEventRef = next;
}
}
bool Event::setStatus(cl_int status) {
int32_t prevStatus = executionStatus;
DBG_LOG(EventsDebugEnable, "setStatus event", this, " new status", status, "previousStatus", prevStatus);
if (isStatusCompleted(&prevStatus)) {
return false;
}
if (status == prevStatus) {
return false;
}
if (peekIsBlocked() && (isStatusCompletedByTermination(&status) == false)) {
return false;
}
if ((status == CL_SUBMITTED) || (isStatusCompleted(&status))) {
bool abortBlockedTasks = isStatusCompletedByTermination(&status);
submitCommand(abortBlockedTasks);
}
this->incRefInternal();
transitionExecutionStatus(status);
if (isStatusCompleted(&status) || (status == CL_SUBMITTED)) {
unblockEventsBlockedByThis(status);
}
executeCallbacks(status);
this->decRefInternal();
return true;
}
void Event::transitionExecutionStatus(int32_t newExecutionStatus) const {
int32_t prevStatus = executionStatus;
DBG_LOG(EventsDebugEnable, "transitionExecutionStatus event", this, " new status", newExecutionStatus, "previousStatus", prevStatus);
while (prevStatus > newExecutionStatus) {
executionStatus.compare_exchange_weak(prevStatus, newExecutionStatus);
}
if (OCLRT::DebugManager.flags.EventsTrackerEnable.get()) {
EventsTracker::getEventsTracker().notifyTransitionedExecutionStatus();
}
}
void Event::submitCommand(bool abortTasks) {
std::unique_ptr<Command> cmdToProcess(cmdToSubmit.exchange(nullptr));
if (cmdToProcess.get() != nullptr) {
if ((this->isProfilingEnabled()) && (this->cmdQueue != nullptr)) {
if (timeStampNode) {
this->cmdQueue->getCommandStreamReceiver().makeResident(*timeStampNode->getGraphicsAllocation());
cmdToProcess->timestamp = timeStampNode->tag;
}
if (profilingCpuPath) {
setSubmitTimeStamp();
setStartTimeStamp();
} else {
this->cmdQueue->getDevice().getOSTime()->getCpuGpuTime(&submitTimeStamp);
}
if (perfCountersEnabled && perfCounterNode) {
this->cmdQueue->getCommandStreamReceiver().makeResident(*perfCounterNode->getGraphicsAllocation());
}
}
auto &complStamp = cmdToProcess->submit(taskLevel, abortTasks);
if (profilingCpuPath && this->isProfilingEnabled() && (this->cmdQueue != nullptr)) {
setEndTimeStamp();
}
updateTaskCount(complStamp.taskCount);
flushStamp->setStamp(complStamp.flushStamp);
submittedCmd.exchange(cmdToProcess.release());
} else if (profilingCpuPath && endTimeStamp == 0) {
setEndTimeStamp();
}
if (this->taskCount == Event::eventNotReady) {
if (!this->isUserEvent() && this->eventWithoutCommand) {
if (this->cmdQueue) {
TakeOwnershipWrapper<Device> deviceOwnerhsip(this->cmdQueue->getDevice());
updateTaskCount(this->cmdQueue->getCommandStreamReceiver().peekTaskCount());
}
}
}
}
cl_int Event::waitForEvents(cl_uint numEvents,
const cl_event *eventList) {
if (numEvents == 0) {
return CL_SUCCESS;
}
//flush all command queues
for (const cl_event *it = eventList, *end = eventList + numEvents; it != end; ++it) {
Event *event = castToObjectOrAbort<Event>(*it);
if (event->cmdQueue) {
if (event->taskLevel != Event::eventNotReady) {
event->cmdQueue->flush();
}
}
}
using WorkerListT = StackVec<cl_event, 64>;
WorkerListT workerList1(eventList, eventList + numEvents);
WorkerListT workerList2;
workerList2.reserve(numEvents);
// pointers to workerLists - for fast swap operations
WorkerListT *currentlyPendingEvents = &workerList1;
WorkerListT *pendingEventsLeft = &workerList2;
while (currentlyPendingEvents->size() > 0) {
for (auto &e : *currentlyPendingEvents) {
Event *event = castToObjectOrAbort<Event>(e);
if (event->peekExecutionStatus() < CL_COMPLETE) {
return CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST;
}
if (event->wait(false, false) == false) {
pendingEventsLeft->push_back(event);
}
}
std::swap(currentlyPendingEvents, pendingEventsLeft);
pendingEventsLeft->clear();
}
return CL_SUCCESS;
}
uint32_t Event::getTaskLevel() {
return taskLevel;
}
inline void Event::unblockEventBy(Event &event, uint32_t taskLevel, int32_t transitionStatus) {
int32_t numEventsBlockingThis = --parentCount;
DEBUG_BREAK_IF(numEventsBlockingThis < 0);
int32_t blockerStatus = transitionStatus;
DEBUG_BREAK_IF(!(isStatusCompleted(&blockerStatus) || peekIsSubmitted(&blockerStatus)));
if ((numEventsBlockingThis > 0) && (isStatusCompletedByTermination(&blockerStatus) == false)) {
return;
}
DBG_LOG(EventsDebugEnable, "Event", this, "is unblocked by", &event);
if (this->taskLevel == Event::eventNotReady) {
this->taskLevel = std::max(cmdQueue->getCommandStreamReceiver().peekTaskLevel(), taskLevel);
} else {
this->taskLevel = std::max(this->taskLevel.load(), taskLevel);
}
int32_t statusToPropagate = CL_SUBMITTED;
if (isStatusCompletedByTermination(&blockerStatus)) {
statusToPropagate = blockerStatus;
}
setStatus(statusToPropagate);
//event may be completed after this operation, transtition the state to not block others.
this->updateExecutionStatus();
}
bool Event::updateStatusAndCheckCompletion() {
auto currentStatus = updateEventAndReturnCurrentStatus();
return isStatusCompleted(&currentStatus);
}
bool Event::isReadyForSubmission() {
return taskLevel != Event::eventNotReady ? true : false;
}
void Event::addCallback(Callback::ClbFuncT fn, cl_int type, void *data) {
ECallbackTarget target = translateToCallbackTarget(type);
if (target == ECallbackTarget::Invalid) {
DEBUG_BREAK_IF(true);
return;
}
incRefInternal();
// Note from spec :
// "All callbacks registered for an event object must be called.
// All enqueued callbacks shall be called before the event object is destroyed."
// That's why each registered calback increments the internal refcount
incRefInternal();
DBG_LOG(EventsDebugEnable, "event", this, "addCallback", "ECallbackTarget", (uint32_t)type);
callbacks[(uint32_t)target].pushFrontOne(*new Callback(this, fn, type, data));
// Callback added after event reached its "completed" state
if (updateStatusAndCheckCompletion()) {
int32_t status = executionStatus;
DBG_LOG(EventsDebugEnable, "event", this, "addCallback executing callbacks with status", status);
executeCallbacks(status);
}
if (peekHasCallbacks() && !isUserEvent() && DebugManager.flags.EnableAsyncEventsHandler.get()) {
platform()->getAsyncEventsHandler()->registerEvent(this);
}
decRefInternal();
}
void Event::executeCallbacks(int32_t executionStatusIn) {
int32_t execStatus = executionStatusIn;
bool terminated = isStatusCompletedByTermination(&execStatus);
ECallbackTarget target;
if (terminated) {
target = ECallbackTarget::Completed;
} else {
target = translateToCallbackTarget(execStatus);
if (target == ECallbackTarget::Invalid) {
DEBUG_BREAK_IF(true);
return;
}
}
// run through all needed callback targets and execute callbacks
for (uint32_t i = 0; i <= (uint32_t)target; ++i) {
auto cb = callbacks[i].detachNodes();
auto curr = cb;
while (curr != nullptr) {
auto next = curr->next;
if (terminated) {
curr->overrideCallbackExecutionStatusTarget(execStatus);
}
DBG_LOG(EventsDebugEnable, "event", this, "executing callback", "ECallbackTarget", (uint32_t)target);
curr->execute();
decRefInternal();
delete curr;
curr = next;
}
}
}
void Event::tryFlushEvent() {
//only if event is not completed, completed event has already been flushed
if (cmdQueue && updateStatusAndCheckCompletion() == false) {
//flush the command queue only if it is not blocked event
if (taskLevel != Event::eventNotReady) {
cmdQueue->getCommandStreamReceiver().flushBatchedSubmissions();
}
}
}
void Event::setQueueTimeStamp() {
if (this->profilingEnabled && (this->cmdQueue != nullptr)) {
this->cmdQueue->getDevice().getOSTime()->getCpuTime(&queueTimeStamp.CPUTimeinNS);
}
}
void Event::setSubmitTimeStamp() {
if (this->profilingEnabled && (this->cmdQueue != nullptr)) {
this->cmdQueue->getDevice().getOSTime()->getCpuTime(&submitTimeStamp.CPUTimeinNS);
}
}
void Event::setStartTimeStamp() {
if (this->profilingEnabled && (this->cmdQueue != nullptr)) {
this->cmdQueue->getDevice().getOSTime()->getCpuTime(&startTimeStamp);
}
}
void Event::setEndTimeStamp() {
if (this->profilingEnabled && (this->cmdQueue != nullptr)) {
this->cmdQueue->getDevice().getOSTime()->getCpuTime(&endTimeStamp);
completeTimeStamp = endTimeStamp;
}
}
TagNode<HwTimeStamps> *Event::getHwTimeStampNode() {
if (!timeStampNode) {
timeStampNode = cmdQueue->getCommandStreamReceiver().getEventTsAllocator()->getTag();
}
return timeStampNode;
}
TagNode<HwPerfCounter> *Event::getHwPerfCounterNode() {
if (!perfCounterNode) {
perfCounterNode = cmdQueue->getCommandStreamReceiver().getEventPerfCountAllocator()->getTag();
}
return perfCounterNode;
}
void Event::copyPerfCounters(InstrPmRegsCfg *config) {
perfConfigurationData = new InstrPmRegsCfg;
memcpy_s(perfConfigurationData, sizeof(InstrPmRegsCfg), config, sizeof(InstrPmRegsCfg));
}
void Event::addTimestampPacketNodes(const TimestampPacketContainer &inputTimestampPacketContainer) {
timestampPacketContainer->assignAndIncrementNodesRefCounts(inputTimestampPacketContainer);
}
TimestampPacketContainer *Event::getTimestampPacketNodes() const { return timestampPacketContainer.get(); }
} // namespace OCLRT