intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-25 21:42:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: read state save area once for all threads for resume(ALL)

Browse Source 
- when resume(all) is called - all threads' sr counter needs to be
verified. Reading state save area separately for all threads takes
longer than reading whole state save area once. State save area is
only read again if sr counter wasn't updated
- fail while reading state save area means threads might have completed
execution
- this fix optimizes time spent in resume(all), that may be called before
debugger detaches

Related-To: NEO-7897

Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>
This commit is contained in:
Mateusz Hoppe
2023-04-14 15:38:30 +00:00
committed by Compute-Runtime-Automation
parent 2cd9c51694
commit bfd32d6284
4 changed files with 300 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

72
level_zero/tools/source/debug/debug_session_imp.cpp
View File

@@ -367,9 +367,10 @@ ze_result_t DebugSessionImp::interrupt(ze_device_thread_t thread) {
return ZE_RESULT_SUCCESS;
}
DebugSessionImp::Error DebugSessionImp::resumeThreadsWithinDevice(uint32_t deviceIndex, ze_device_thread_t physicalThread) {
DebugSessionImp::Error DebugSessionImp::resumeThreadsWithinDevice(uint32_t deviceIndex, ze_device_thread_t apiThread) {
auto &hwInfo = connectedDevice->getHwInfo();
bool allThreadsRunning = true;
auto physicalThread = convertToPhysicalWithinDevice(apiThread, deviceIndex);
auto singleThreads = getSingleThreadsForDevice(deviceIndex, physicalThread, hwInfo);
Error retVal = Error::Unknown;
uint64_t memoryHandle = EuThread::invalidHandle;
@@ -400,13 +401,42 @@ DebugSessionImp::Error DebugSessionImp::resumeThreadsWithinDevice(uint32_t devic
auto result = resumeImp(resumeThreadIds, deviceIndex);
for (auto &threadID : resumeThreadIds) {
while (checkThreadIsResumed(threadID) == false)
;
// For resume(ALL) and multiple threads to resume - read whole state save area
// to avoid multiple calls to KMD
if (resumeThreadIds.size() > 1 && isThreadAll(apiThread)) {
allThreads[threadID]->resumeThread();
auto gpuVa = getContextStateSaveAreaGpuVa(memoryHandle);
auto stateSaveAreaSize = getContextStateSaveAreaSize(memoryHandle);
std::unique_ptr<char[]> stateSaveArea = nullptr;
auto stateSaveReadResult = ZE_RESULT_ERROR_UNKNOWN;
if (gpuVa != 0 && stateSaveAreaSize != 0) {
stateSaveArea = std::make_unique<char[]>(stateSaveAreaSize);
stateSaveReadResult = readGpuMemory(memoryHandle, stateSaveArea.get(), stateSaveAreaSize, gpuVa);
} else {
DEBUG_BREAK_IF(true);
}
for (auto &threadID : resumeThreadIds) {
if (stateSaveReadResult == ZE_RESULT_SUCCESS) {
while (checkThreadIsResumed(threadID, stateSaveArea.get()) == false) {
// read state save area again to update sr counters
stateSaveReadResult = readGpuMemory(allThreads[threadID]->getMemoryHandle(), stateSaveArea.get(), stateSaveAreaSize, gpuVa);
}
}
allThreads[threadID]->resumeThread();
}
} else {
for (auto &threadID : resumeThreadIds) {
while (checkThreadIsResumed(threadID) == false)
;
allThreads[threadID]->resumeThread();
}
}
checkStoppedThreadsAndGenerateEvents(resumeThreadIds, memoryHandle, deviceIndex);
if (sipCommandResult && result == ZE_RESULT_SUCCESS) {
@@ -506,6 +536,28 @@ bool DebugSessionImp::checkThreadIsResumed(const EuThread::ThreadId &threadID) {
return resumed;
}
bool DebugSessionImp::checkThreadIsResumed(const EuThread::ThreadId &threadID, const void *stateSaveArea) {
auto stateSaveAreaHeader = getStateSaveAreaHeader();
bool resumed = true;
if (stateSaveAreaHeader->versionHeader.version.major >= 2u) {
SIP::sr_ident srMagic = {{0}};
const auto thread = allThreads[threadID].get();
if (!readSystemRoutineIdentFromMemory(thread, stateSaveArea, srMagic)) {
return resumed;
}
PRINT_DEBUGGER_THREAD_LOG("checkThreadIsResumed - Read counter for thread %s, counter == %d\n", EuThread::toString(threadID).c_str(), (int)srMagic.count);
// Counter greater than last one means thread was resumed
if (srMagic.count == thread->getLastCounter()) {
resumed = false;
}
}
return resumed;
}
ze_result_t DebugSessionImp::resume(ze_device_thread_t thread) {
auto deviceCount = std::max(1u, connectedDevice->getNEODevice()->getNumSubDevices());
@@ -522,8 +574,8 @@ ze_result_t DebugSessionImp::resume(ze_device_thread_t thread) {
deviceIndex = getDeviceIndexFromApiThread(thread);
}
}
auto physicalThread = convertToPhysicalWithinDevice(thread, deviceIndex);
auto result = resumeThreadsWithinDevice(deviceIndex, physicalThread);
auto result = resumeThreadsWithinDevice(deviceIndex, thread);
if (result == Error::ThreadsRunning) {
retVal = ZE_RESULT_ERROR_NOT_AVAILABLE;
@@ -534,9 +586,7 @@ ze_result_t DebugSessionImp::resume(ze_device_thread_t thread) {
bool allThreadsRunning = true;
for (uint32_t deviceId = 0; deviceId < deviceCount; deviceId++) {
auto physicalThread = convertToPhysicalWithinDevice(thread, deviceId);
auto result = resumeThreadsWithinDevice(deviceId, physicalThread);
auto result = resumeThreadsWithinDevice(deviceId, thread);
if (result == Error::ThreadsRunning) {
continue;