intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-09 06:23:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: avoid overflow of gpu time stamp in ns

Browse Source 
Related-To: NEO-8394
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski
2023-10-24 08:11:23 +00:00
committed by Compute-Runtime-Automation
parent 1fd9d29f54
commit 6d2d16d68e
7 changed files with 32 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
level_zero/core/source/event/event_impl.inl
View File

@@ -610,8 +610,8 @@ void EventImp<TagSizeT>::getSynchronizedKernelTimestamps(ze_synchronized_timesta
auto &gfxCoreHelper = device->getNEODevice()->getGfxCoreHelper();
auto &hwInfo = device->getNEODevice()->getHardwareInfo();
const auto frequency = device->getNEODevice()->getDeviceInfo().profilingTimerResolution;
auto deviceTsInNs = gfxCoreHelper.getGpuTimeStampInNS(referenceTs.gpuTimeStamp, frequency);
const auto resolution = device->getNEODevice()->getDeviceInfo().profilingTimerResolution;
auto deviceTsInNs = gfxCoreHelper.getGpuTimeStampInNS(referenceTs.gpuTimeStamp, resolution);
const auto maxKernelTsValue = maxNBitValue(hwInfo.capabilityTable.kernelTimestampValidBits);
auto getDuration = [&](uint64_t startTs, uint64_t endTs) {
@@ -640,15 +640,15 @@ void EventImp<TagSizeT>::getSynchronizedKernelTimestamps(ze_synchronized_timesta
auto calculateSynchronizedTs = [&](ze_synchronized_timestamp_data_ext_t *synchronizedTs, const ze_kernel_timestamp_data_t *deviceTs) {
// Add the offset to the kernel timestamp to find the start timestamp on the CPU timescale
int64_t offset = tsOffsetInNs;
uint64_t startTimeStampInNs = static_cast<uint64_t>(deviceTs->kernelStart * frequency) + offset;
uint64_t startTimeStampInNs = static_cast<uint64_t>(deviceTs->kernelStart * resolution) + offset;
if (startTimeStampInNs < referenceHostTsInNs) {
offset += static_cast<uint64_t>(maxNBitValue(gfxCoreHelper.getGlobalTimeStampBits()) * frequency);
startTimeStampInNs = static_cast<uint64_t>(deviceTs->kernelStart * frequency) + offset;
offset += static_cast<uint64_t>(maxNBitValue(gfxCoreHelper.getGlobalTimeStampBits()) * resolution);
startTimeStampInNs = static_cast<uint64_t>(deviceTs->kernelStart * resolution) + offset;
}
// Get the kernel timestamp duration
uint64_t deviceDuration = getDuration(deviceTs->kernelStart, deviceTs->kernelEnd);
uint64_t deviceDurationNs = static_cast<uint64_t>(deviceDuration * frequency);
uint64_t deviceDurationNs = static_cast<uint64_t>(deviceDuration * resolution);
// Add the duration to the startTimeStamp to get the endTimeStamp
uint64_t endTimeStampInNs = startTimeStampInNs + deviceDurationNs;
@@ -662,7 +662,7 @@ void EventImp<TagSizeT>::getSynchronizedKernelTimestamps(ze_synchronized_timesta
pSynchronizedTimestampsBuffer[index].context.kernelStart = pSynchronizedTimestampsBuffer[index].global.kernelStart;
uint64_t deviceDuration = getDuration(pKernelTimestampsBuffer[index].context.kernelStart,
pKernelTimestampsBuffer[index].context.kernelEnd);
uint64_t deviceDurationNs = static_cast<uint64_t>(deviceDuration * frequency);
uint64_t deviceDurationNs = static_cast<uint64_t>(deviceDuration * resolution);
pSynchronizedTimestampsBuffer[index].context.kernelEnd = pSynchronizedTimestampsBuffer[index].context.kernelStart +
deviceDurationNs;
}