intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-09 06:23:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: skip Events from previous append in in-order CL path

Browse Source 
Related-To: NEO-7966

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
2023-06-01 09:57:35 +00:00
committed by Compute-Runtime-Automation
parent 808ff8c2e4
commit fb7fff7442
3 changed files with 74 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -2176,6 +2176,15 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWaitOnEvents(uint32_t nu
continue;
}
if (isInOrderExecutionEnabled() && event->isInOrderExecEvent()) {
auto const cmdListNode = this->timestampPacketContainer->peekNodes()[0];
auto const eventNode = event->getInOrderTimestampPacket()->peekNodes()[0];
if (cmdListNode == eventNode) {
continue;
}
}
commandContainer.addToResidencyContainer(&event->getAllocation(this->device));
gpuAddr = event->getCompletionFieldGpuAddress(this->device);
uint32_t packetsToWait = event->getPacketsInUse();

2
level_zero/core/source/event/event.h
View File

@@ -205,6 +205,8 @@ struct Event : _ze_event_handle_t {
this->metricStreamer = metricStreamer;
}
void enableInOrderExecMode(const NEO::TimestampPacketContainer &inOrderSyncNodes);
bool isInOrderExecEvent() const { return inOrderExecEvent; }
const NEO::TimestampPacketContainer *getInOrderTimestampPacket() const { return inOrderTimestampPacket.get(); }
protected:
Event(EventPool *eventPool, int index, Device *device) : device(device), eventPool(eventPool), index(index) {}

63
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_3.cpp
View File

@@ -806,6 +806,69 @@ HWTEST2_F(InOrderCmdListTests, givenInOrderModeWhenSubmittingThenProgramSemaphor
EXPECT_EQ(MI_SEMAPHORE_WAIT::COMPARE_OPERATION::COMPARE_OPERATION_SAD_NOT_EQUAL_SDD, semaphoreCmd->getCompareOperation());
}
HWTEST2_F(InOrderCmdListTests, givenInOrderModeWhenWaitingForEventFromPreviousAppendThenSkip, IsAtLeastXeHpCore) {
using MI_SEMAPHORE_WAIT = typename FamilyType::MI_SEMAPHORE_WAIT;
auto immCmdList = createImmCmdList<gfxCoreFamily>();
auto eventPool = createEvents<FamilyType>(1, false);
auto eventHandle = events[0]->toHandle();
auto cmdStream = immCmdList->getCmdContainer().getCommandStream();
immCmdList->appendLaunchKernel(kernel->toHandle(), &groupCount, eventHandle, 0, nullptr, launchParams, false);
auto offset = cmdStream->getUsed();
immCmdList->appendLaunchKernel(kernel->toHandle(), &groupCount, nullptr, 1, &eventHandle, launchParams, false);
GenCmdList cmdList;
ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
cmdList,
ptrOffset(cmdStream->getCpuBase(), offset),
cmdStream->getUsed() - offset));
auto itor = find<typename FamilyType::MI_SEMAPHORE_WAIT *>(cmdList.begin(), cmdList.end());
ASSERT_NE(cmdList.end(), itor); // implicit dependency
itor = find<typename FamilyType::MI_SEMAPHORE_WAIT *>(++itor, cmdList.end());
EXPECT_EQ(cmdList.end(), itor);
}
HWTEST2_F(InOrderCmdListTests, givenInOrderModeWhenWaitingForEventFromAfterResetThenDontSkip, IsAtLeastXeHpCore) {
using MI_SEMAPHORE_WAIT = typename FamilyType::MI_SEMAPHORE_WAIT;
auto immCmdList = createImmCmdList<gfxCoreFamily>();
auto eventPool = createEvents<FamilyType>(1, false);
auto eventHandle = events[0]->toHandle();
auto cmdStream = immCmdList->getCmdContainer().getCommandStream();
immCmdList->appendLaunchKernel(kernel->toHandle(), &groupCount, eventHandle, 0, nullptr, launchParams, false);
events[0]->reset();
auto offset = cmdStream->getUsed();
immCmdList->appendLaunchKernel(kernel->toHandle(), &groupCount, nullptr, 1, &eventHandle, launchParams, false);
GenCmdList cmdList;
ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
cmdList,
ptrOffset(cmdStream->getCpuBase(), offset),
cmdStream->getUsed() - offset));
auto itor = find<typename FamilyType::MI_SEMAPHORE_WAIT *>(cmdList.begin(), cmdList.end());
ASSERT_NE(cmdList.end(), itor); // implicit dependency
itor = find<typename FamilyType::MI_SEMAPHORE_WAIT *>(++itor, cmdList.end());
EXPECT_NE(cmdList.end(), itor);
}
HWTEST2_F(InOrderCmdListTests, givenInOrderEventModeWhenSubmittingThenProgramSemaphoreForEvent, IsAtLeastXeHpCore) {
using MI_SEMAPHORE_WAIT = typename FamilyType::MI_SEMAPHORE_WAIT;