intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-04 15:53:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: correct sequence of estimates to get correct size for start command

Browse Source 
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2024-11-26 14:06:41 +00:00
committed by Compute-Runtime-Automation
parent 974e8ae63f
commit 6453a5ec31
4 changed files with 92 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -329,8 +329,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushImmediateTask(
handleImmediateFlushStateBaseAddressState(dispatchFlags, flushData, device);
handleImmediateFlushOneTimeContextInitState(dispatchFlags, flushData, device);
handleImmediateFlushJumpToImmediate(flushData);
bool stateCacheFlushRequired = device.getBindlessHeapsHelper() ? device.getBindlessHeapsHelper()->getStateDirtyForContext(getOsContext().getContextId()) : false;
if (stateCacheFlushRequired) {
flushData.estimatedSize += MemorySynchronizationCommands<GfxFamily>::getSizeForFullCacheFlush();
@@ -340,6 +338,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushImmediateTask(
flushData.estimatedSize += MemorySynchronizationCommands<GfxFamily>::getSizeForInstructionCacheFlush();
}
// this must be the last call after all estimate size operations
handleImmediateFlushJumpToImmediate(flushData);
auto &csrCommandStream = getCS(flushData.estimatedSize);
flushData.csrStartOffset = csrCommandStream.getUsed();

4
shared/test/common/fixtures/command_stream_receiver_fixture.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2023 Intel Corporation
* Copyright (C) 2022-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -17,7 +17,7 @@ void CommandStreamReceiverFixture::setUp() {
DeviceFixture::setUp();
commandStream.replaceBuffer(cmdBuffer, bufferSize);
auto graphicsAllocation = new MockGraphicsAllocation(cmdBuffer, bufferSize);
auto graphicsAllocation = new MockGraphicsAllocation(cmdBuffer, cmdBufferGpuAddress, bufferSize);
commandStream.replaceGraphicsAllocation(graphicsAllocation);
dsh.replaceBuffer(dshBuffer, bufferSize);

3
shared/test/common/fixtures/command_stream_receiver_fixture.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022-2023 Intel Corporation
* Copyright (C) 2022-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -17,6 +17,7 @@ struct CommandStreamReceiverFixture : public NEO::DeviceFixture {
void setUp();
void tearDown();
static constexpr uint64_t cmdBufferGpuAddress = 0x10000;
static constexpr size_t bufferSize = 1024;
uint8_t cmdBuffer[bufferSize];
uint8_t dshBuffer[bufferSize];

89
shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
View File

@@ -5111,8 +5111,6 @@ HWTEST_F(CommandStreamReceiverHwTest, GivenDirtyFlagForContextInBindlessHelperWh
bindlessHeapsHelperPtr->stateCacheDirtyForContext.set(commandStreamReceiver.getOsContext().getContextId());
this->requiredStreamProperties.stateComputeMode.setPropertiesAll(false, GrfConfig::defaultGrfNumber, ThreadArbitrationPolicy::AgeBased, NEO::PreemptionMode::ThreadGroup);
commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
HardwareParse hwParserCsr;
@@ -5128,6 +5126,51 @@ HWTEST_F(CommandStreamReceiverHwTest, GivenDirtyFlagForContextInBindlessHelperWh
EXPECT_FALSE(bindlessHeapsHelperPtr->getStateDirtyForContext(commandStreamReceiver.getOsContext().getContextId()));
}
HWTEST_F(CommandStreamReceiverHwTest, GivenContextInitializedAndDirtyFlagForContextInBindlessHelperWhenFlushImmediateTaskCalledThenStateCacheInvalidateIsSentBeforeBbStartJumpingToImmediateBuffer) {
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
// 1st flush, dispatch all required n-p state commands
commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
auto usedAfterFirstFlush = commandStreamReceiver.commandStream.getUsed();
auto bindlessHeapsHelper = std::make_unique<MockBindlesHeapsHelper>(pDevice, pDevice->getNumGenericSubDevices() > 1);
MockBindlesHeapsHelper *bindlessHeapsHelperPtr = bindlessHeapsHelper.get();
pDevice->getExecutionEnvironment()->rootDeviceEnvironments[pDevice->getRootDeviceIndex()]->bindlessHeapsHelper.reset(bindlessHeapsHelper.release());
bindlessHeapsHelperPtr->stateCacheDirtyForContext.set(commandStreamReceiver.getOsContext().getContextId());
// only state cache flush is dispatched in dynamic preamble
auto immediateBufferStartOffset = commandStream.getUsed();
commandStreamReceiver.flushImmediateTask(commandStream, immediateBufferStartOffset, immediateFlushTaskFlags, *pDevice);
HardwareParse hwParserCsr;
hwParserCsr.parseCommands<FamilyType>(commandStreamReceiver.commandStream, usedAfterFirstFlush);
auto cmdPtrIt = hwParserCsr.cmdList.begin();
ASSERT_NE(hwParserCsr.cmdList.end(), cmdPtrIt);
auto pipeControl = genCmdCast<PIPE_CONTROL *>(*cmdPtrIt);
ASSERT_NE(nullptr, pipeControl);
EXPECT_TRUE(pipeControl->getCommandStreamerStallEnable());
EXPECT_TRUE(pipeControl->getStateCacheInvalidationEnable());
EXPECT_TRUE(pipeControl->getTextureCacheInvalidationEnable());
EXPECT_TRUE(pipeControl->getRenderTargetCacheFlushEnable());
EXPECT_FALSE(bindlessHeapsHelperPtr->getStateDirtyForContext(commandStreamReceiver.getOsContext().getContextId()));
cmdPtrIt++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdPtrIt);
auto bbStart = genCmdCast<MI_BATCH_BUFFER_START *>(*cmdPtrIt);
ASSERT_NE(nullptr, bbStart);
EXPECT_EQ(cmdBufferGpuAddress + immediateBufferStartOffset, bbStart->getBatchBufferStartAddress());
}
HWTEST_F(CommandStreamReceiverHwTest, givenRequiresInstructionCacheFlushWhenFlushImmediateThenInstructionCacheInvalidateEnableIsSent) {
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
@@ -5137,8 +5180,6 @@ HWTEST_F(CommandStreamReceiverHwTest, givenRequiresInstructionCacheFlushWhenFlus
}
commandStreamReceiver.registerInstructionCacheFlush();
this->requiredStreamProperties.stateComputeMode.setPropertiesAll(false, GrfConfig::defaultGrfNumber, ThreadArbitrationPolicy::AgeBased, NEO::PreemptionMode::ThreadGroup);
commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
HardwareParse hwParserCsr;
@@ -5150,6 +5191,46 @@ HWTEST_F(CommandStreamReceiverHwTest, givenRequiresInstructionCacheFlushWhenFlus
EXPECT_FALSE(commandStreamReceiver.requiresInstructionCacheFlush);
}
HWTEST_F(CommandStreamReceiverHwTest, givenContextInitializedAndRequiresInstructionCacheFlushWhenFlushImmediateThenInstructionCacheInvalidateEnableIsSentBeforeBbStartJumpingToImmediateBuffer) {
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
if (commandStreamReceiver.heaplessStateInitialized) {
GTEST_SKIP();
}
// 1st flush, dispatch all required n-p state commands
commandStreamReceiver.flushImmediateTask(commandStream, commandStream.getUsed(), immediateFlushTaskFlags, *pDevice);
auto usedAfterFirstFlush = commandStreamReceiver.commandStream.getUsed();
commandStreamReceiver.registerInstructionCacheFlush();
// only instruction cache flush is dispatched in dynamic preamble
auto immediateBufferStartOffset = commandStream.getUsed();
commandStreamReceiver.flushImmediateTask(commandStream, immediateBufferStartOffset, immediateFlushTaskFlags, *pDevice);
HardwareParse hwParserCsr;
hwParserCsr.parseCommands<FamilyType>(commandStreamReceiver.commandStream, usedAfterFirstFlush);
auto cmdPtrIt = hwParserCsr.cmdList.begin();
ASSERT_NE(hwParserCsr.cmdList.end(), cmdPtrIt);
auto pipeControl = genCmdCast<PIPE_CONTROL *>(*cmdPtrIt);
ASSERT_NE(nullptr, pipeControl);
EXPECT_TRUE(pipeControl->getInstructionCacheInvalidateEnable());
EXPECT_FALSE(commandStreamReceiver.requiresInstructionCacheFlush);
cmdPtrIt++;
ASSERT_NE(hwParserCsr.cmdList.end(), cmdPtrIt);
auto bbStart = genCmdCast<MI_BATCH_BUFFER_START *>(*cmdPtrIt);
ASSERT_NE(nullptr, bbStart);
EXPECT_EQ(cmdBufferGpuAddress + immediateBufferStartOffset, bbStart->getBatchBufferStartAddress());
}
HWTEST_F(CommandStreamReceiverHwTest, GivenFlushIsBlockingWhenFlushTaskCalledThenExpectMonitorFenceFlagTrue) {
auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
commandStreamReceiver.recordFlushedBatchBuffer = true;