vsw
October 9, 2023, 8:38pm
1
Hi,
How to convert frame timestamp to system time. AGX has gps receiver with pps.
I read all topics but it is not clear for me.
Now i am here:
EGLStream::IArgusCaptureMetadata * iArgusCaptureMetadata{ Argus::interface_cast<EGLStream::IArgusCaptureMetadata >( frame ) };
Argus::CaptureMetadata * metadata{ iArgusCaptureMetadata->getMetadata() };
Argus::Ext::ISensorTimestampTsc * iSensorTimestampTsc{ Argus::interface_cast< Argus::Ext::ISensorTimestampTsc >( metadata ) };
const auto ts{ iSensorTimestampTsc->getSensorSofTimestampTsc() - m_offsetNs };
const auto monotonicRawNow{ hc::utils::getMonotonicRawNow() }; // clock_gettime
const auto nanoNow{ hc::utils::nano() }; // chrono now
const auto tsNow{ nanoNow - ( monotonicRawNow - ts ) };
tsNow is real clock time?
Best regards, Viktor.
Suppose the monotonicRawNow and nanoNow could have time gap due to it get in different calls.
vsw
October 11, 2023, 12:35pm
4
Is it right formula?
( monotonicRawNow - ts ) always about 1ms.
Nano is RTC time.
How to obtain frame RTC time?
Here is my funcs:
uint64_t getOffsetNs()
{
std::ifstream buf{ "/sys/devices/system/clocksource/clocksource0/offset_ns" };
std::stringstream strStream;
strStream << buf.rdbuf();
return std::stoull( strStream.str() );
}
uint64_t getMonotonicRawNow()
{
// Get the current time from steady_clock
struct timespec tss;
clock_gettime( CLOCK_MONOTONIC_RAW, &tss );
return { uint64_t( 1000000000UL ) * int64_t( tss.tv_sec ) + int64_t( tss.tv_nsec ) };
}
Please help.
Using the timspec64 and ns_to_timespec64 would be good.
vsw
October 11, 2023, 2:53pm
6
ShaneCCC:
ns_to_timespec64
Yes, of course, for 64bit arch timespec used 64bit digits.
vsw
October 11, 2023, 2:58pm
7
So the question about:
epoch_ns = X + StartOfFrame_ns
How to find X?
What’s the epoch_ns time?
vsw
October 11, 2023, 3:04pm
9
System time from 1970 in nanoseconds
I don’t think there’s a way to get it precisely.
vsw
October 11, 2023, 3:11pm
11
What the best precision (gap) can i achieve? Which way?
Please advice me, i will try.
vsw
October 11, 2023, 7:15pm
12
Here is my approach:
Stored monotonic_raw ( startTime ) and now ( startTs ) when first frame acquired.
time = startTime + ( ts - startTs )
Am i right?
#pragma omp parallel for schedule(static) num_threads(6)
for( cameraId = 0; cameraId < m_camerasNum; ++ cameraId )
{
Argus::UniqueObj< EGLStream::Frame > frame{ m_iFrameConsumers.at( cameraId )->acquireFrame() };
EGLStream::IArgusCaptureMetadata * iArgusCaptureMetadata{ Argus::interface_cast< EGLStream::IArgusCaptureMetadata >( frame ) };
Argus::CaptureMetadata * metadata{ iArgusCaptureMetadata->getMetadata() };
Argus::Ext::ISensorTimestampTsc * iSensorTimestampTsc{ Argus::interface_cast< Argus::Ext::ISensorTimestampTsc >( metadata ) };
EGLStream::IFrame * iFrame{ Argus::interface_cast< EGLStream::IFrame >( frame ) };
if( not iFrame )
{
std::cerr << "Failed to acquire frame! " << cameraId << std::endl;
} else
{
const auto ts{ iSensorTimestampTsc->getSensorSofTimestampTsc() - m_offsetNs };
const auto monotonicRawNow{ hc::utils::getMonotonicRawNow() };
const auto nanoNow{ hc::utils::nano() };
const auto diff{ monotonicRawNow - ts };
if( startTs.at( cameraId ) == 0 and startTime.at( cameraId ) == 0 )
{
startTs.at( cameraId ) = monotonicRawNow;
startTime.at( cameraId ) = nanoNow;
}
const auto time{ startTime.at( cameraId ) + ( ts - startTs.at( cameraId ) ) };
const auto timeStr{
"Timestamp: " +
std::to_string( cameraId ) + " " +
std::to_string( iFrame->getNumber() ) + " " +
std::to_string( ts / 1000000 ) + " " +
std::to_string( time / 1000000 ) + " " +
std::to_string( diff / 1000000 ) + " " +
std::to_string( nanoNow / 1000000 )
};
// record
if( m_doCapture and iFrame->getNumber() > m_skipFrames )
{
std::cout << timeStr << std::endl;
EGLStream::NV::IImageNativeBuffer * iNativeBuffer{ Argus::interface_cast< EGLStream::NV::IImageNativeBuffer >( iFrame->getImage()) };
if( not iNativeBuffer )
std::cerr << "IImageNativeBuffer not supported by Image!" << std::endl;
else
{
if( m_dmaBuffers.at( cameraId ) == - 1 )
m_dmaBuffers.at( cameraId ) = iNativeBuffer->createNvBuffer( m_streamSize, NvBufferColorFormat_YUV420, NvBufferLayout_Pitch );
if( m_dmaBuffers.at( cameraId ) != - 1 )
{
if( iNativeBuffer->copyToNvBuffer( m_dmaBuffers.at( cameraId ) ) != Argus::STATUS_OK )
{
std::cerr << "Failed to copy frame to NvBuffer!" << std::endl;
} else
{
const auto filename{ m_recordPath + "/" + std::to_string( cameraId ) + "/" + std::to_string( iFrame->getNumber() ) + ".jpg" };
std::ofstream outputFile{ filename };
if( outputFile.is_open() )
{
uint8_t * buffer{ m_OutputBuffers.at( cameraId ) };
unsigned long size{ bufferSize };
if( m_JpegEncoders.at( cameraId )->encodeFromFd( m_dmaBuffers.at( cameraId ), JCS_YCbCr, & buffer, size ) == 0 )
{
outputFile.write( ( char * ) buffer, size );
dataFile << timeStr << std::endl;
} else
std::cout << "Failed to encode jpeg!" << std::endl;
}
}
} else
std::cerr << "Failed to create NvBuffer!" << std::endl;
}
}
}
}
This is hardware synchronized 6 sensors imx264 24fps with saving frames to ssd. Where every frame has GPS time ( gps receiver with pps connected to AGX ).
Here is result in milliseconds:
Timestamp: 4 224 21475137 1697055720376 47 1697055720424
vsw
October 13, 2023, 1:18pm
13
Do i need to subtract offsetNs from sensor timestamp in JP4.6.1?
Both of them are need to substract offsetNs to convert to host system time.
vsw
October 20, 2023, 10:46am
16
Here i made more calculations.
iMetadata->getSensorTimestamp() do not need subtract offset. It already has it.
Timestamp: 0 147 ts:1088121 ts2:1078565 ts3:1088169 of:9603
Timestamp: 5 126 ts:1088121 ts2:1078565 ts3:1088169 of:9603
Timestamp: 1 147 ts:1088121 ts2:1078565 ts3:1088169 of:9603
Timestamp: 3 147 ts:1088121 ts2:1078565 ts3:1088169 of:9603
Timestamp: 2 148 ts:1088161 ts2:1078606 ts3:1088209 of:9603
Timestamp: 4 148 ts:1088161 ts2:1078606 ts3:1088209 of:9603
Timestamp: 0 148 ts:1088161 ts2:1078606 ts3:1088209 of:9603
ts - getSensorSofTimestampTsc() - m_offsetNs
Please comment it.
So getSensorSofTimestampTsc() - getSensorTimestamp() has offset?
They are start of frame time both, are not they?
Please help.
I think the getSensorSofTimestampTsc() is the same with getSensorTimestamp()
vsw
October 20, 2023, 3:21pm
18
But they never equal. Why? Did you read my calculatons?
vsw
October 20, 2023, 3:42pm
19
EGLStream::IArgusCaptureMetadata * iArgusCaptureMetadata{ Argus::interface_cast< EGLStream::IArgusCaptureMetadata >( frame ) };
Argus::CaptureMetadata * metadata{ iArgusCaptureMetadata->getMetadata() };
Argus::Ext::ISensorTimestampTsc * iSensorTimestampTsc{ Argus::interface_cast< Argus::Ext::ISensorTimestampTsc >( metadata ) };
const Argus::ICaptureMetadata * iMetadata{ Argus::interface_cast< const Argus::ICaptureMetadata >( metadata ) };
const auto ts{ iSensorTimestampTsc->getSensorSofTimestampTsc() - offsetNs
const auto ts3{ iMetadata->getSensorTimestamp() };
|ts - ts3| ~ 50-500ms
It is right code?
vsw
October 20, 2023, 3:59pm
20
Can you provide data which shows that you think?
As i see in JP4.6.1:
getSensorTimestamp() do not need to subtract offset_ns
getSensorSofTimestampTsc() and getSensorTimestamp() very differs
May be it depends on hardware?
Can you point me to any post topic where issue was investigated and confirmed?
vsw
October 20, 2023, 4:31pm
21
May be Nvidia have real developer to help consumers?
ShaneCCC are you developer?
