intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 23:03:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Use single event for multiple kernels

Browse Source 
Related-To: NEO-6871

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2022-04-08 18:48:45 +00:00
committed by Compute-Runtime-Automation
parent a3745c28a3
commit 819d648997
25 changed files with 1074 additions and 338 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
level_zero/core/source/event/event.h
View File

@@ -96,6 +96,17 @@ struct Event : _ze_event_handle_t {
return isTimestampEvent || usingContextEndOffset;
}
void increaseKernelCount() {
kernelCount++;
UNRECOVERABLE_IF(kernelCount > EventPacketsCount::maxKernelSplit);
}
uint32_t getKernelCount() const {
return kernelCount;
}
void zeroKernelCount() {
kernelCount = 0;
}
uint64_t globalStartTS;
uint64_t globalEndTS;
uint64_t contextStartTS;
@@ -110,8 +121,6 @@ struct Event : _ze_event_handle_t {
ze_event_scope_flags_t signalScope = 0u;
ze_event_scope_flags_t waitScope = 0u;
uint32_t kernelCount = 1u;
bool l3FlushWaApplied = false;
protected:
@@ -122,6 +131,9 @@ struct Event : _ze_event_handle_t {
size_t timestampSizeInDw = 0u;
size_t singlePacketSize = 0u;
size_t eventPoolOffset = 0u;
uint32_t kernelCount = 1u;
bool isTimestampEvent = false;
bool usingContextEndOffset = false;
};
@@ -180,8 +192,7 @@ struct EventImp : public Event {
protected:
ze_result_t calculateProfilingData();
ze_result_t queryStatusKernelTimestamp();
ze_result_t queryStatusNonTimestamp();
ze_result_t queryStatusEventPackets();
ze_result_t hostEventSetValue(TagSizeT eventValue);
ze_result_t hostEventSetValueTimestamps(TagSizeT eventVal);
void assignKernelEventCompletionData(void *address);

38
level_zero/core/source/event/event_impl.inl
View File

@@ -104,33 +104,13 @@ void EventImp<TagSizeT>::assignKernelEventCompletionData(void *address) {
}
template <typename TagSizeT>
ze_result_t EventImp<TagSizeT>::queryStatusKernelTimestamp() {
ze_result_t EventImp<TagSizeT>::queryStatusEventPackets() {
assignKernelEventCompletionData(hostAddress);
uint32_t queryVal = Event::STATE_CLEARED;
for (uint32_t i = 0; i < kernelCount; i++) {
uint32_t packetsToCheck = kernelEventCompletionData[i].getPacketsUsed();
for (uint32_t packetId = 0; packetId < packetsToCheck; packetId++) {
bool ready = NEO::WaitUtils::waitFunctionWithPredicate<const TagSizeT>(
static_cast<TagSizeT const *>(kernelEventCompletionData[i].getContextEndAddress(packetId)),
queryVal,
std::not_equal_to<TagSizeT>());
if (!ready) {
return ZE_RESULT_NOT_READY;
}
}
}
this->csr->getInternalAllocationStorage()->cleanAllocationList(this->csr->peekTaskCount(), NEO::AllocationUsage::TEMPORARY_ALLOCATION);
return ZE_RESULT_SUCCESS;
}
template <typename TagSizeT>
ze_result_t EventImp<TagSizeT>::queryStatusNonTimestamp() {
assignKernelEventCompletionData(hostAddress);
uint32_t queryVal = Event::STATE_CLEARED;
for (uint32_t i = 0; i < kernelCount; i++) {
uint32_t packetsToCheck = kernelEventCompletionData[i].getPacketsUsed();
for (uint32_t packetId = 0; packetId < packetsToCheck; packetId++) {
void const *queryAddress = usingContextEndOffset
void const *queryAddress = isUsingContextEndOffset()
? kernelEventCompletionData[i].getContextEndAddress(packetId)
: kernelEventCompletionData[i].getContextStartAddress(packetId);
bool ready = NEO::WaitUtils::waitFunctionWithPredicate<const TagSizeT>(
@@ -156,11 +136,7 @@ ze_result_t EventImp<TagSizeT>::queryStatus() {
*hostAddr = metricStreamer->getNotificationState();
}
this->csr->downloadAllocations();
if (isEventTimestampFlagSet()) {
return queryStatusKernelTimestamp();
} else {
return queryStatusNonTimestamp();
}
return queryStatusEventPackets();
}
template <typename TagSizeT>
@@ -274,11 +250,9 @@ ze_result_t EventImp<TagSizeT>::hostSynchronize(uint64_t timeout) {
template <typename TagSizeT>
ze_result_t EventImp<TagSizeT>::reset() {
if (isEventTimestampFlagSet()) {
kernelCount = EventPacketsCount::maxKernelSplit;
for (uint32_t i = 0; i < kernelCount; i++) {
kernelEventCompletionData[i].setPacketsUsed(NEO::TimestampPacketSizeControl::preferredPacketCount);
}
kernelCount = EventPacketsCount::maxKernelSplit;
for (uint32_t i = 0; i < kernelCount; i++) {
kernelEventCompletionData[i].setPacketsUsed(NEO::TimestampPacketSizeControl::preferredPacketCount);
}
hostEventSetValue(Event::STATE_INITIAL);
resetPackets();