intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-08 22:12:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "performance(ocl): program barrier pc in taskStream"

Browse Source 
This reverts commit 839c2d6737.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
This commit is contained in:
Compute-Runtime-Validation
2023-09-12 01:11:41 +02:00
committed by Compute-Runtime-Automation
parent 0a28d622fa
commit b5e9c10f64
16 changed files with 40 additions and 163 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
opencl/source/command_queue/command_queue.h
View File

@@ -348,12 +348,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
void fillCsrDependenciesWithLastBcsPackets(CsrDependencies &csrDeps);
void clearLastBcsPackets();
void setStallingCommandsOnNextFlush(const bool isStallingCommandsOnNextFlushRequired) {
stallingCommandsOnNextFlushRequired = isStallingCommandsOnNextFlushRequired;
if (!isStallingCommandsOnNextFlushRequired) {
dcFlushRequiredOnStallingCommandsOnNextFlush = false;
}
}
void setStallingCommandsOnNextFlush(const bool isStallingCommandsOnNextFlushRequired) { stallingCommandsOnNextFlushRequired = isStallingCommandsOnNextFlushRequired; }
bool isStallingCommandsOnNextFlushRequired() const { return stallingCommandsOnNextFlushRequired; }
void setDcFlushRequiredOnStallingCommandsOnNextFlush(const bool isDcFlushRequiredOnStallingCommandsOnNextFlush) { dcFlushRequiredOnStallingCommandsOnNextFlush = isDcFlushRequiredOnStallingCommandsOnNextFlush; }

21
opencl/source/command_queue/enqueue_common.h
View File

@@ -272,10 +272,11 @@ cl_int CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
} else if (computeCommandStreamReceiver.peekTimestampPacketWriteEnabled()) {
if (CL_COMMAND_BARRIER == commandType && !isNonStallingIoqBarrier) {
setStallingCommandsOnNextFlush(true);
const bool isDcFlushRequiredOnBarrier = NEO::DebugManager.flags.SkipDcFlushOnBarrierWithoutEvents.get() == 0 || event;
setDcFlushRequiredOnStallingCommandsOnNextFlush(isDcFlushRequiredOnBarrier);
if (NEO::DebugManager.flags.SkipDcFlushOnBarrierWithoutEvents.get() == 0 || event) {
setDcFlushRequiredOnStallingCommandsOnNextFlush(true);
}
this->splitBarrierRequired = true;
}
this->splitBarrierRequired = true;
for (size_t i = 0; i < eventsRequest.numEventsInWaitList; i++) {
auto waitlistEvent = castToObjectOrAbort<Event>(eventsRequest.eventWaitList[i]);
@@ -777,18 +778,6 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
UNRECOVERABLE_IF(multiDispatchInfo.empty());
if (!relaxedOrderingEnabled && !getGpgpuCommandStreamReceiver().isMultiTileOperationEnabled() && isStallingCommandsOnNextFlushRequired() && !isBlitAuxTranslationRequired(multiDispatchInfo)) {
CsrDependencies csrDeps{};
fillCsrDependenciesWithLastBcsPackets(csrDeps);
TimestampPacketHelper::programCsrDependenciesForTimestampPacketContainer<GfxFamily>(commandStream, csrDeps, false);
setupBarrierTimestampForBcsEngines(getGpgpuCommandStreamReceiver().getOsContext().getEngineType(), timestampPacketDependencies);
getGpgpuCommandStreamReceiver().programStallingCommandsForBarrier(commandStream, &timestampPacketDependencies.barrierNodes, isDcFlushRequiredOnStallingCommandsOnNextFlush());
clearLastBcsPackets();
setStallingCommandsOnNextFlush(false);
}
auto implicitFlush = false;
if (printfHandler) {
@@ -975,6 +964,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
if (isHandlingBarrier) {
clearLastBcsPackets();
setStallingCommandsOnNextFlush(false);
setDcFlushRequiredOnStallingCommandsOnNextFlush(false);
}
if (gtpinIsGTPinInitialized()) {
@@ -1195,6 +1185,7 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueCommandWithoutKernel(
if (isHandlingBarrier) {
clearLastBcsPackets();
setStallingCommandsOnNextFlush(false);
setDcFlushRequiredOnStallingCommandsOnNextFlush(false);
}
}

2
opencl/source/helpers/task_information.cpp
View File

@@ -273,6 +273,7 @@ CompletionStamp &CommandComputeKernel::submit(TaskCountType taskLevel, bool term
if (isHandlingBarrier) {
commandQueue.clearLastBcsPackets();
commandQueue.setStallingCommandsOnNextFlush(false);
commandQueue.setDcFlushRequiredOnStallingCommandsOnNextFlush(false);
}
if (kernelOperation->blitPropertiesContainer.size() > 0) {
@@ -433,6 +434,7 @@ CompletionStamp &CommandWithoutKernel::submit(TaskCountType taskLevel, bool term
if (isHandlingBarrier) {
commandQueue.clearLastBcsPackets();
commandQueue.setStallingCommandsOnNextFlush(false);
commandQueue.setDcFlushRequiredOnStallingCommandsOnNextFlush(false);
}
if (kernelOperation->blitEnqueue) {

16
opencl/test/unit_test/command_queue/command_queue_hw_2_tests.cpp
View File

@@ -1186,20 +1186,23 @@ HWTEST_F(OoqCommandQueueHwBlitTest, givenBlitBeforeBarrierWhenEnqueueingCommandT
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
uint64_t lastBlitNodeAddress = TimestampPacketHelper::getContextEndGpuAddress(*pCmdQ->getTimestampPacketContainer()->peekNodes()[0]);
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueKernel(kernel, 1, &offset, &gws, nullptr, 0, nullptr, nullptr));
auto ccsStart = pCmdQ->getGpgpuCommandStreamReceiver().getCS().getUsed();
auto bcsStart = pCmdQ->getBcsCommandStreamReceiver(aub_stream::ENGINE_BCS)->getCS(0).getUsed();
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueBarrierWithWaitList(0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueKernel(kernel, 1, &offset, &gws, nullptr, 0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
uint64_t barrierNodeAddress = 0u;
{
HardwareParse queueHwParser;
queueHwParser.parseCommands<FamilyType>(*pDevice->getUltCommandStreamReceiver<FamilyType>().lastFlushedCommandStream, 0);
HardwareParse ccsHwParser;
ccsHwParser.parseCommands<FamilyType>(pCmdQ->getGpgpuCommandStreamReceiver().getCS(0), ccsStart);
const auto semaphoreItor = find<MI_SEMAPHORE_WAIT *>(queueHwParser.cmdList.begin(), queueHwParser.cmdList.end());
const auto semaphoreItor = find<MI_SEMAPHORE_WAIT *>(ccsHwParser.cmdList.begin(), ccsHwParser.cmdList.end());
const auto semaphore = genCmdCast<MI_SEMAPHORE_WAIT *>(*semaphoreItor);
EXPECT_EQ(lastBlitNodeAddress, semaphore->getSemaphoreGraphicsAddress());
const auto pipeControlItor = find<PIPE_CONTROL *>(semaphoreItor, queueHwParser.cmdList.end());
const auto pipeControlItor = find<PIPE_CONTROL *>(semaphoreItor, ccsHwParser.cmdList.end());
const auto pipeControl = genCmdCast<PIPE_CONTROL *>(*pipeControlItor);
EXPECT_EQ(PIPE_CONTROL::POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA, pipeControl->getPostSyncOperation());
barrierNodeAddress = pipeControl->getAddress() | (static_cast<uint64_t>(pipeControl->getAddressHigh()) << 32);
@@ -1208,9 +1211,6 @@ HWTEST_F(OoqCommandQueueHwBlitTest, givenBlitBeforeBarrierWhenEnqueueingCommandT
EXPECT_EQ(pipeControlItor, find<MI_SEMAPHORE_WAIT *>(std::next(semaphoreItor), pipeControlItor));
}
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
{
HardwareParse bcsHwParser;
bcsHwParser.parseCommands<FamilyType>(pCmdQ->getBcsCommandStreamReceiver(aub_stream::ENGINE_BCS)->getCS(0), bcsStart);

10
opencl/test/unit_test/command_queue/command_queue_tests_pvc_and_later.cpp
View File

@@ -820,18 +820,16 @@ HWTEST2_F(OoqCommandQueueHwBlitTest, givenBarrierBeforeFirstKernelWhenEnqueueNDR
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueReadBuffer(buffer.get(), CL_FALSE, 0, 1u, ptr, nullptr, 0, nullptr, nullptr));
auto ccsStart = pCmdQ->getGpgpuCommandStreamReceiver().getCS().getUsed();
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueBarrierWithWaitList(0, nullptr, nullptr));
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueKernel(kernel, 1, &offset, &gws, nullptr, 0, nullptr, nullptr));
auto ccsStart = pCmdQ->getGpgpuCommandStreamReceiver().getCS().getUsed();
HardwareParse queueHwParser;
queueHwParser.parseCommands<FamilyType>(*pDevice->getUltCommandStreamReceiver<FamilyType>().lastFlushedCommandStream, 0u);
const auto memFenceItor = find<MI_MEM_FENCE *>(queueHwParser.cmdList.begin(), queueHwParser.cmdList.end());
EXPECT_NE(queueHwParser.cmdList.end(), memFenceItor);
EXPECT_EQ(CL_SUCCESS, pCmdQ->enqueueKernel(kernel, 1, &offset, &gws, nullptr, 0, nullptr, nullptr));
HardwareParse ccsHwParser;
ccsHwParser.parseCommands<FamilyType>(pCmdQ->getGpgpuCommandStreamReceiver().getCS(0), ccsStart);
const auto memFenceStateItor = find<STATE_SYSTEM_MEM_FENCE_ADDRESS *>(ccsHwParser.cmdList.begin(), ccsHwParser.cmdList.end());
const auto memFenceItor = find<MI_MEM_FENCE *>(memFenceStateItor, ccsHwParser.cmdList.end());
EXPECT_NE(ccsHwParser.cmdList.end(), memFenceItor);
EXPECT_NE(ccsHwParser.cmdList.end(), memFenceStateItor);
}

10
opencl/test/unit_test/command_queue/enqueue_barrier_tests.cpp
View File

@@ -26,9 +26,7 @@ using BarrierTest = Test<CommandEnqueueFixture>;
HWTEST_F(BarrierTest, givenCsrWithHigherLevelThenCommandQueueWhenEnqueueBarrierIsCalledThenCommandQueueAlignsToCsrWithoutSendingAnyCommands) {
auto pCmdQ = this->pCmdQ;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
GTEST_SKIP();
}
// Set task levels to known values.
uint32_t originalCSRLevel = 2;
commandStreamReceiver.taskLevel = originalCSRLevel;
@@ -71,9 +69,7 @@ HWTEST_F(BarrierTest, GivenCsrTaskLevelGreaterThenCmdqTaskLevelWhenEnqueingBarri
auto pCmdQ = this->pCmdQ;
auto pCmdBuffer = this->pCmdBuffer;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
GTEST_SKIP();
}
commandStreamReceiver.setMediaVFEStateDirty(false);
// Set task levels to known values.
@@ -206,7 +202,7 @@ HWTEST_F(BarrierTest, WhenEnqueingBarrierWithWaitListThenDependenciesShouldSync)
auto pEvent = castToObject<Event>(event);
auto &csr = pCmdQ->getGpgpuCommandStreamReceiver();
// in this case only cmdQ raises the taskLevel while csr stays intact
// in this case only cmdQ raises the taskLevel why csr stay intact
EXPECT_EQ(8u, pCmdQ->taskLevel);
if (csr.peekTimestampPacketWriteEnabled()) {
EXPECT_EQ(pCmdQ->taskLevel + 1, commandStreamReceiver.peekTaskLevel());

6
opencl/test/unit_test/command_queue/ooq_task_tests.cpp
View File

@@ -5,9 +5,7 @@
*
*/
#include "shared/test/common/cmd_parse/hw_parse.h"
#include "shared/test/common/helpers/debug_manager_state_restore.h"
#include "shared/test/common/helpers/unit_test_helper.h"
#include "shared/test/common/mocks/mock_csr.h"
#include "shared/test/common/test_macros/test_checks_shared.h"
#include "shared/test/common/utilities/base_object_utils.h"
@@ -370,7 +368,7 @@ HWTEST_F(OOQTaskTests, givenSkipDcFlushOnBarrierWithEventsEnabledWhenEnqueingBar
outEvent->release();
}
HWTEST_F(OOQTaskTests, givenSkipDcFlushOnBarrierWithoutEventsDisabledWhenEnqueingBarrierWithWaitListThenDcFlushSet) {
HWTEST_F(OOQTaskTests, givenSkipDcFlushOnBarrierWithoutEventsDisableddWhenEnqueingBarrierWithWaitListThenDcFlushSet) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (false == commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
GTEST_SKIP();
@@ -390,7 +388,7 @@ HWTEST_F(OOQTaskTests, givenSkipDcFlushOnBarrierWithoutEventsDisabledWhenEnquein
EXPECT_TRUE(pCmdQ->isDcFlushRequiredOnStallingCommandsOnNextFlush());
}
HWTEST_F(OOQTaskTests, givenEnqueueMarkerWithWaitListWhenIsMarkerWithPostSyncWriteThenBcsTimestampLastBarrierToWaitForIsNotEmpty) {
HWTEST_F(OOQTaskTests, givenEnqueueMarkerWithWaitListWhenIsMarkerWithPostSyncWriteThenBcsTimestapLastBarrierToWaitForIsNotEmpty) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (false == commandStreamReceiver.peekTimestampPacketWriteEnabled()) {
GTEST_SKIP();

2
opencl/test/unit_test/command_stream/command_stream_receiver_hw_1_tests.cpp
View File

@@ -1755,7 +1755,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, UltCommandStreamReceiverTest, givenBarrierNodeSetWhe
size_t estimatedCmdSize = commandStreamReceiver->getCmdSizeForStallingCommands(dispatchFlags);
EXPECT_EQ(expectedCmdSize, estimatedCmdSize);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags.barrierTimestampPacketNodes, false);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags);
EXPECT_EQ(estimatedCmdSize, commandStreamCSR.getUsed());
parseCommands<FamilyType>(commandStreamCSR, 0);

6
opencl/test/unit_test/command_stream/command_stream_receiver_hw_tests_xehp_and_later.cpp
View File

@@ -945,7 +945,7 @@ HWTEST2_F(CommandStreamReceiverHwTestXeHPAndLater, givenStaticPartitionEnabledWh
size_t estimatedCmdSize = commandStreamReceiver->getCmdSizeForStallingCommands(dispatchFlags);
EXPECT_EQ(expectedCmdSize, estimatedCmdSize);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags.barrierTimestampPacketNodes, false);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags);
EXPECT_EQ(estimatedCmdSize, commandStreamCSR.getUsed());
parseCommands<FamilyType>(commandStreamCSR, 0);
@@ -992,7 +992,7 @@ HWTEST2_F(CommandStreamReceiverHwTestXeHPAndLater, givenStaticPartitionDisabledW
size_t estimatedCmdSize = commandStreamReceiver->getCmdSizeForStallingCommands(dispatchFlags);
EXPECT_EQ(expectedCmdSize, estimatedCmdSize);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags.barrierTimestampPacketNodes, false);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags);
EXPECT_EQ(estimatedCmdSize, commandStreamCSR.getUsed());
parseCommands<FamilyType>(commandStreamCSR, 0);
@@ -1045,7 +1045,7 @@ HWTEST2_F(CommandStreamReceiverHwTestXeHPAndLater, givenStaticPartitionEnabledWh
size_t estimatedCmdSize = commandStreamReceiver->getCmdSizeForStallingCommands(dispatchFlags);
EXPECT_EQ(expectedSize, estimatedCmdSize);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags.barrierTimestampPacketNodes, false);
commandStreamReceiver->programStallingCommandsForBarrier(commandStreamCSR, dispatchFlags);
EXPECT_EQ(estimatedCmdSize, commandStreamCSR.getUsed());
EXPECT_EQ(2u, tagNode->getPacketsUsed());

2
opencl/test/unit_test/kernel/kernel_tests.cpp
View File

@@ -591,8 +591,6 @@ class CommandStreamReceiverMock : public CommandStreamReceiver {
size_t getCmdsSizeForComputeBarrierCommand() const override {
return 0;
}
void programStallingCommandsForBarrier(LinearStream &cmdStream, TimestampPacketContainer *barrierTimestampPacketNodes, const bool isDcFlushRequired) override {
}
GraphicsAllocation *getClearColorAllocation() override { return nullptr; }
bool createPreemptionAllocation() override {

4
opencl/test/unit_test/mem_obj/buffer_bcs_tests.cpp
View File

@@ -811,7 +811,7 @@ HWTEST_TEMPLATED_F(BcsBufferTests, givenPipeControlRequestWhenDispatchingBlitEnq
}
}
HWTEST_TEMPLATED_F(BcsBufferTests, givenStallingCommandsOnNextFlushWhenReleasingMultipleBlockedEnqueuesThenProgramBarrierOnce) {
HWTEST_TEMPLATED_F(BcsBufferTests, givenBarrierWithEmptyWaitlistWhenReleasingMultipleBlockedEnqueuesThenProgramBarrierOnce) {
DebugManager.flags.OptimizeIoqBarriersHandling.set(0);
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
@@ -827,7 +827,7 @@ HWTEST_TEMPLATED_F(BcsBufferTests, givenStallingCommandsOnNextFlushWhenReleasing
cl_event waitlist0[] = {&userEvent0};
cl_event waitlist1[] = {&userEvent1};
cmdQ->setStallingCommandsOnNextFlush(true);
cmdQ->enqueueBarrierWithWaitList(0, nullptr, nullptr);
cmdQ->enqueueWriteBuffer(buffer.get(), false, 0, 1, hostPtr, nullptr, 1, waitlist0, nullptr);
cmdQ->enqueueWriteBuffer(buffer.get(), false, 0, 1, hostPtr, nullptr, 1, waitlist1, nullptr);