intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-20 17:00:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Do not set dcFlush in Immediate dispatch mode.

Browse Source 
Some devices do not need dcFlush.
Setting it prevents further optimization of pipe controls which
are not needed.

Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek
2022-03-02 12:43:02 +00:00
committed by Compute-Runtime-Automation
parent bfacd14b61
commit 8e94d568a8
2 changed files with 50 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_2_tests.cpp
View File

@@ -80,6 +80,50 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenBlockedKernelNotRequiringDCFl
buffer->release();
}
HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenEnableUpdateTaskFromWaitWhenNonBlockingCallIsMadeThenNoPipeControlInsertedOnDevicesWithoutDCFlushRequirements) {
DebugManagerStateRestore restorer;
DebugManager.flags.UpdateTaskCountFromWait.set(3u);
typedef typename FamilyType::PIPE_CONTROL PIPE_CONTROL;
MockContext ctx(pClDevice);
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
commandStreamReceiver.timestampPacketWriteEnabled = false;
CommandQueueHw<FamilyType> commandQueue(&ctx, pClDevice, 0, false);
size_t tempBuffer[] = {0, 1, 2};
size_t dstBuffer[] = {0, 1, 2};
cl_int retVal = 0;
auto buffer = Buffer::create(&ctx, CL_MEM_USE_HOST_PTR, sizeof(tempBuffer), tempBuffer, retVal);
commandQueue.enqueueWriteBuffer(buffer, CL_FALSE, 0, sizeof(tempBuffer), dstBuffer, nullptr, 0u, nullptr, 0);
auto &commandStreamTask = *commandStreamReceiver.lastFlushedCommandStream;
cmdList.clear();
// Parse command list
parseCommands<FamilyType>(commandStreamTask, 0);
auto pipeControlExpected = MemorySynchronizationCommands<FamilyType>::getDcFlushEnable(true, *defaultHwInfo);
auto itorPC = find<PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
if (pipeControlExpected) {
EXPECT_NE(cmdList.end(), itorPC);
if (UnitTestHelper<FamilyType>::isPipeControlWArequired(pDevice->getHardwareInfo())) {
itorPC++;
itorPC = find<PIPE_CONTROL *>(itorPC, cmdList.end());
EXPECT_NE(cmdList.end(), itorPC);
}
// Verify that the dcFlushEnabled bit is set in PC
auto pCmdWA = reinterpret_cast<PIPE_CONTROL *>(*itorPC);
EXPECT_EQ(MemorySynchronizationCommands<FamilyType>::getDcFlushEnable(true, *defaultHwInfo), pCmdWA->getDcFlushEnable());
} else {
EXPECT_EQ(cmdList.end(), itorPC);
}
buffer->release();
}
HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenTaskCsPassedAsCommandStreamParamWhenFlushingTaskThenCompletionStampIsCorrect) {
CommandQueueHw<FamilyType> commandQueue(nullptr, pClDevice, 0, false);
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();

9
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -187,6 +187,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
}
const auto &hwInfo = peekHwInfo();
auto &hwHelper = HwHelper::get(hwInfo.platform.eRenderCoreFamily);
bool updateTag = false;
if (dispatchFlags.blocking || dispatchFlags.dcFlush || dispatchFlags.guardCommandBufferWithPipeControl) {
if (this->dispatchMode == DispatchMode::ImmediateDispatch) {
@@ -195,7 +197,10 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
levelClosed = true;
//if we guard with ppc, flush dc as well to speed up completion latency
if (dispatchFlags.guardCommandBufferWithPipeControl) {
dispatchFlags.dcFlush = true;
const auto &hwInfoConfig = *NEO::HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isDcFlushAllowed()) {
dispatchFlags.dcFlush = true;
}
}
}
@@ -255,8 +260,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
dispatchFlags.pipelineSelectArgs.specialPipelineSelectMode,
hwInfo);
auto &hwHelper = HwHelper::get(peekHwInfo().platform.eRenderCoreFamily);
if (dispatchFlags.threadArbitrationPolicy == ThreadArbitrationPolicy::NotPresent) {
if (this->streamProperties.stateComputeMode.threadArbitrationPolicy.value != -1) {
// Reuse previous programming