intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-22 10:17:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix for incorrect timestamp offset calculation in event profiling info

Browse Source 
Change-Id: I634c29daf4734b24e4075542dc6550c531977f0a
Signed-off-by: Pawel Wilma <pawel.wilma@intel.com>
This commit is contained in:
Pawel Wilma
2019-01-09 17:07:11 +01:00
committed by sys_ocldev
parent 397778df70
commit 14e8fdd8f8
3 changed files with 91 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
runtime/command_queue/gpgpu_walker.inl
View File

@@ -106,12 +106,12 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
// PIPE_CONTROL for global timestamp
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->GlobalStartTS, hwTimeStamps.tag);
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->GlobalStartTS, hwTimeStamps.getGraphicsAllocation()->getUnderlyingBuffer());
PipeControlHelper<GfxFamily>::obtainPipeControlAndProgramPostSyncOperation(commandStream, PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP, TimeStampAddress, 0llu);
//MI_STORE_REGISTER_MEM for context local timestamp
TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextStartTS, hwTimeStamps.tag);
TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextStartTS, hwTimeStamps.getGraphicsAllocation()->getUnderlyingBuffer());
//low part
auto pMICmdLow = (MI_STORE_REGISTER_MEM *)commandStream->getSpace(sizeof(MI_STORE_REGISTER_MEM));
@@ -134,7 +134,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsEnd(
pPipeControlCmd->setCommandStreamerStallEnable(true);
//MI_STORE_REGISTER_MEM for context local timestamp
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextEndTS, hwTimeStamps.tag);
uint64_t TimeStampAddress = hwTimeStamps.getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamps.tag->ContextEndTS, hwTimeStamps.getGraphicsAllocation()->getUnderlyingBuffer());
//low part
auto pMICmdLow = (MI_STORE_REGISTER_MEM *)commandStream->getSpace(sizeof(MI_STORE_REGISTER_MEM));

4
runtime/device_queue/device_queue_hw.inl
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2018 Intel Corporation
* Copyright (C) 2017-2019 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -216,7 +216,7 @@ void DeviceQueueHw<GfxFamily>::addExecutionModelCleanUpSection(Kernel *parentKer
using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
if (hwTimeStamp != nullptr) {
uint64_t TimeStampAddress = hwTimeStamp->getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamp->tag->ContextCompleteTS, hwTimeStamp->tag);
uint64_t TimeStampAddress = hwTimeStamp->getGraphicsAllocation()->getGpuAddress() + ptrDiff(&hwTimeStamp->tag->ContextCompleteTS, hwTimeStamp->getGraphicsAllocation()->getUnderlyingBuffer());
igilQueue->m_controls.m_EventTimestampAddress = TimeStampAddress;
addProfilingEndCmds(TimeStampAddress);

87
unit_tests/command_queue/dispatch_walker_tests.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2017-2018 Intel Corporation
* Copyright (C) 2017-2019 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -12,6 +12,7 @@
#include "runtime/helpers/aligned_memory.h"
#include "runtime/helpers/kernel_commands.h"
#include "runtime/helpers/task_information.h"
#include "runtime/utilities/tag_allocator.h"
#include "unit_tests/fixtures/device_fixture.h"
#include "unit_tests/command_queue/command_queue_fixture.h"
#include "unit_tests/libult/mock_gfx_family.h"
@@ -1143,3 +1144,87 @@ HWTEST_F(DispatchWalkerTest, givenKernelWhenAuxTranslationWithoutParentKernelThe
auto itorCmd = find<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
ASSERT_NE(cmdList.end(), itorCmd);
}
struct ProfilingCommandsTest : public DispatchWalkerTest, ::testing::WithParamInterface<bool> {
void SetUp() override {
DispatchWalkerTest::SetUp();
}
void TearDown() override {
DispatchWalkerTest::TearDown();
}
};
HWTEST_P(ProfilingCommandsTest, givenKernelWhenProfilingCommandStartIsTakenThenTimeStampAddressIsProgrammedCorrectly) {
using MI_STORE_REGISTER_MEM = typename FamilyType::MI_STORE_REGISTER_MEM;
using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
bool checkForStart = GetParam();
auto &cmdStream = pCmdQ->getCS(0);
TagAllocator<HwTimeStamps> timeStampAllocator(this->pDevice->getMemoryManager(), 10, MemoryConstants::cacheLineSize);
auto hwTimeStamp1 = timeStampAllocator.getTag();
ASSERT_NE(nullptr, hwTimeStamp1);
if (checkForStart) {
GpgpuWalkerHelper<FamilyType>::dispatchProfilingCommandsStart(*hwTimeStamp1, &cmdStream);
} else {
GpgpuWalkerHelper<FamilyType>::dispatchProfilingCommandsEnd(*hwTimeStamp1, &cmdStream);
}
auto hwTimeStamp2 = timeStampAllocator.getTag();
ASSERT_NE(nullptr, hwTimeStamp2);
if (checkForStart) {
GpgpuWalkerHelper<FamilyType>::dispatchProfilingCommandsStart(*hwTimeStamp2, &cmdStream);
} else {
GpgpuWalkerHelper<FamilyType>::dispatchProfilingCommandsEnd(*hwTimeStamp2, &cmdStream);
}
GenCmdList cmdList;
ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(cmdList, cmdStream.getCpuBase(), cmdStream.getUsed()));
auto itorStoreReg = find<typename FamilyType::MI_STORE_REGISTER_MEM *>(cmdList.begin(), cmdList.end());
ASSERT_NE(cmdList.end(), itorStoreReg);
auto storeReg = genCmdCast<MI_STORE_REGISTER_MEM *>(*itorStoreReg);
ASSERT_NE(nullptr, storeReg);
uint64_t gpuAddress = storeReg->getMemoryAddress();
auto timestampFieldAddress = checkForStart ? &hwTimeStamp1->tag->ContextStartTS : &hwTimeStamp1->tag->ContextEndTS;
uint64_t expectedAddress = hwTimeStamp1->getGraphicsAllocation()->getGpuAddress() + ptrDiff(timestampFieldAddress, hwTimeStamp1->getGraphicsAllocation()->getUnderlyingBuffer());
EXPECT_EQ(expectedAddress, gpuAddress);
itorStoreReg++;
itorStoreReg = find<typename FamilyType::MI_STORE_REGISTER_MEM *>(itorStoreReg, cmdList.end());
ASSERT_NE(cmdList.end(), itorStoreReg);
storeReg = genCmdCast<MI_STORE_REGISTER_MEM *>(*itorStoreReg);
ASSERT_NE(nullptr, storeReg);
gpuAddress = storeReg->getMemoryAddress();
timestampFieldAddress = checkForStart ? &hwTimeStamp2->tag->ContextStartTS : &hwTimeStamp2->tag->ContextEndTS;
expectedAddress = hwTimeStamp2->getGraphicsAllocation()->getGpuAddress() + ptrDiff(timestampFieldAddress, hwTimeStamp2->getGraphicsAllocation()->getUnderlyingBuffer());
EXPECT_EQ(expectedAddress, gpuAddress);
if (checkForStart) {
auto itorPipeCtrl = find<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
ASSERT_NE(cmdList.end(), itorPipeCtrl);
auto pipeControl = genCmdCast<PIPE_CONTROL *>(*itorPipeCtrl);
ASSERT_NE(nullptr, pipeControl);
gpuAddress = static_cast<uint64_t>(pipeControl->getAddress()) | (static_cast<uint64_t>(pipeControl->getAddressHigh()) << 32);
timestampFieldAddress = checkForStart ? &hwTimeStamp1->tag->GlobalStartTS : &hwTimeStamp1->tag->GlobalEndTS;
expectedAddress = hwTimeStamp1->getGraphicsAllocation()->getGpuAddress() + ptrDiff(timestampFieldAddress, hwTimeStamp1->getGraphicsAllocation()->getUnderlyingBuffer());
EXPECT_EQ(expectedAddress, gpuAddress);
itorPipeCtrl++;
itorPipeCtrl = find<typename FamilyType::PIPE_CONTROL *>(itorPipeCtrl, cmdList.end());
ASSERT_NE(cmdList.end(), itorPipeCtrl);
pipeControl = genCmdCast<PIPE_CONTROL *>(*itorPipeCtrl);
ASSERT_NE(nullptr, pipeControl);
gpuAddress = static_cast<uint64_t>(pipeControl->getAddress()) | static_cast<uint64_t>(pipeControl->getAddressHigh()) << 32;
timestampFieldAddress = checkForStart ? &hwTimeStamp2->tag->GlobalStartTS : &hwTimeStamp2->tag->GlobalEndTS;
expectedAddress = hwTimeStamp2->getGraphicsAllocation()->getGpuAddress() + ptrDiff(timestampFieldAddress, hwTimeStamp2->getGraphicsAllocation()->getUnderlyingBuffer());
EXPECT_EQ(expectedAddress, gpuAddress);
}
}
INSTANTIATE_TEST_CASE_P(StartEndFlag,
ProfilingCommandsTest, ::testing::Values(true, false));