intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-06 19:32:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Complete kernel prior to timestamp capture

Browse Source 
While at it, fix event flag check upon creation

Change-Id: I2c57b0e20141fc38a4de695cae79bf2f14dd6cdd
Signed-off-by: Aravind Gopalakrishnan <Aravind.Gopalakrishnan@intel.com>
This commit is contained in:
Aravind Gopalakrishnan
2020-04-08 12:44:14 -07:00
committed by sys_ocldev
parent bd24350a68
commit 71ff85cf2c
6 changed files with 224 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
level_zero/core/source/cmdlist/cmdlist_hw_base.inl
View File

@@ -92,17 +92,12 @@ void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfiling(ze_event_hand
commandContainer.addToResidencyContainer(&event->getAllocation());
if (beforeWalker) {
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::GLOBAL_START_LOW);
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::GLOBAL_START);
NEO::EncodeStoreMMIO<GfxFamily>::encode(commandContainer, REG_GLOBAL_TIMESTAMP_LDW, timeStampAddress);
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::GLOBAL_START_HIGH);
NEO::EncodeStoreMMIO<GfxFamily>::encode(commandContainer, REG_GLOBAL_TIMESTAMP_UN, timeStampAddress);
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::CONTEXT_START);
NEO::EncodeStoreMMIO<GfxFamily>::encode(commandContainer, GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW, timeStampAddress);
} else {
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::CONTEXT_END);
NEO::EncodeStoreMMIO<GfxFamily>::encode(commandContainer, GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW, timeStampAddress);
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::GLOBAL_END);
bool dcFlushEnable = (event->signalScope == ZE_EVENT_SCOPE_FLAG_NONE) ? false : true;
@@ -116,6 +111,13 @@ void CommandListCoreFamily<gfxCoreFamily>::appendEventForProfiling(ze_event_hand
0llu,
dcFlushEnable,
device->getHwInfo());
timeStampAddress = event->getGpuAddress() + event->getOffsetOfEventTimestampRegister(Event::CONTEXT_END);
NEO::EncodeStoreMMIO<GfxFamily>::encode(commandContainer, GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW, timeStampAddress);
if (dcFlushEnable) {
NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControl(*commandContainer.getCommandStream(), true);
}
}
}
}

25
level_zero/core/source/event/event.cpp
View File

@@ -67,8 +67,8 @@ struct EventImp : public Event {
EventPool *eventPool;
protected:
ze_result_t hostEventSetValue(uint64_t eventValue);
ze_result_t hostEventSetValueTimestamps(uint64_t eventVal);
ze_result_t hostEventSetValue(uint32_t eventValue);
ze_result_t hostEventSetValueTimestamps(uint32_t eventVal);
void makeAllocationResident();
};
@@ -81,7 +81,7 @@ struct EventPoolImp : public EventPool {
}
auto timestampMultiplier = 1;
if (flags == ZE_EVENT_POOL_FLAG_TIMESTAMP) {
if (flags & ZE_EVENT_POOL_FLAG_TIMESTAMP) {
isEventPoolUsedForTimestamp = true;
timestampMultiplier = numEventTimestampsToRead;
}
@@ -154,7 +154,7 @@ struct EventPoolImp : public EventPool {
protected:
const uint32_t eventSize = 16u;
const uint32_t eventAlignment = MemoryConstants::cacheLineSize;
const int32_t numEventTimestampsToRead = 5u;
const int32_t numEventTimestampsToRead = 4u;
};
Event *Event::create(EventPool *eventPool, const ze_event_desc_t *desc, Device *device) {
@@ -206,7 +206,7 @@ void EventImp::makeAllocationResident() {
}
}
ze_result_t EventImp::hostEventSetValueTimestamps(uint64_t eventVal) {
ze_result_t EventImp::hostEventSetValueTimestamps(uint32_t eventVal) {
for (uint32_t i = 0; i < this->eventPool->getNumEventTimestampsToRead(); i++) {
auto baseAddr = reinterpret_cast<uint64_t>(hostAddress);
auto timeStampAddress = baseAddr + getOffsetOfEventTimestampRegister(i);
@@ -224,7 +224,7 @@ ze_result_t EventImp::hostEventSetValueTimestamps(uint64_t eventVal) {
return ZE_RESULT_SUCCESS;
}
ze_result_t EventImp::hostEventSetValue(uint64_t eventVal) {
ze_result_t EventImp::hostEventSetValue(uint32_t eventVal) {
if (isTimestampEvent) {
hostEventSetValueTimestamps(eventVal);
}
@@ -292,8 +292,8 @@ ze_result_t EventImp::reset() {
ze_result_t EventImp::getTimestamp(ze_event_timestamp_type_t timestampType, void *dstptr) {
auto baseAddr = reinterpret_cast<uint64_t>(hostAddress);
uint64_t *tsptr = nullptr;
uint64_t tsData = Event::STATE_INITIAL;
constexpr uint64_t tsMask = (1ull << 32) - 1;
uint64_t tsData = Event::STATE_INITIAL & tsMask;
if (!this->isTimestampEvent)
return ZE_RESULT_ERROR_INVALID_ARGUMENT;
@@ -305,15 +305,8 @@ ze_result_t EventImp::getTimestamp(ze_event_timestamp_type_t timestampType, void
}
if (timestampType == ZE_EVENT_TIMESTAMP_GLOBAL_START) {
tsptr = reinterpret_cast<uint64_t *>(baseAddr + getOffsetOfEventTimestampRegister(Event::GLOBAL_START_LOW));
auto tsptrUpper = reinterpret_cast<uint64_t *>(baseAddr + getOffsetOfEventTimestampRegister(Event::GLOBAL_START_HIGH));
tsData = ((*tsptrUpper & tsMask) << 32) | (*tsptr & tsMask);
memcpy_s(dstptr, sizeof(uint64_t), static_cast<void *>(&tsData), sizeof(uint64_t));
return ZE_RESULT_SUCCESS;
}
if (timestampType == ZE_EVENT_TIMESTAMP_GLOBAL_END) {
tsptr = reinterpret_cast<uint64_t *>(baseAddr + getOffsetOfEventTimestampRegister(Event::GLOBAL_START));
} else if (timestampType == ZE_EVENT_TIMESTAMP_GLOBAL_END) {
tsptr = reinterpret_cast<uint64_t *>(baseAddr + getOffsetOfEventTimestampRegister(Event::GLOBAL_END));
} else if (timestampType == ZE_EVENT_TIMESTAMP_CONTEXT_START) {
tsptr = reinterpret_cast<uint64_t *>(baseAddr + getOffsetOfEventTimestampRegister(Event::CONTEXT_START));

7
level_zero/core/source/event/event.h
View File

@@ -30,15 +30,14 @@ struct Event : _ze_event_handle_t {
virtual ze_result_t reset() = 0;
virtual ze_result_t getTimestamp(ze_event_timestamp_type_t timestampType, void *dstptr) = 0;
enum State : uint64_t {
enum State : uint32_t {
STATE_SIGNALED = 0u,
STATE_CLEARED = static_cast<uint64_t>(-1),
STATE_CLEARED = static_cast<uint32_t>(-1),
STATE_INITIAL = STATE_CLEARED
};
enum EventTimestampRegister : uint32_t {
GLOBAL_START_LOW = 0u,
GLOBAL_START_HIGH,
GLOBAL_START = 0u,
GLOBAL_END,
CONTEXT_START,
CONTEXT_END