Davide,

Have you allocated enough memory for the TX buffer (are you sure, that the hard fault is not because you address an to out-of-RAM region)?
Are you aware of the restrictions related to the RAIL's TX FIFO?

The transmit FIFO must also be byte-aligned, with no further restrictions. Its size has the same restrictions as the receive FIFO. The size must be a power of 2 from 64 to 4096 for hardware compatibility.

If it doesn't help, I would like to get a minimal reproducer program to debug it on my side: the .sls format would be the best exported by Simplicity Studio.

Best Regards,
Bence

Andrew,

You wrote 20 byte, but I think it's only a typo, and you meant bit, however it doesn't matter, your calculation seems correct. But I want to highlight it again: this configuration is not a typical LDC use case, since the radio will spend only the fifth of the time in SLEEP state, and I don't think it is worth enough to save so little amount of energy (it's about ~50% comparing to the case when the radio is always in RX, but it's still very high if e.g. the radio powered by a coin cell).

And I should repeat myself regarding the timing in practice: even if the calculations are correct you must measure both TX and RX timings to check whether those work as expected (and adjust them if not).

Best Regards,
Bence

Hi Bas,

I'm almost sure that 800 bit preamble isn't enough for 26 channels, as it means that every channel have ~31 bit time including the channel switching to signal a valid preamble in your current solution. I suggest to measure the turnaround time (enable RAIL_EVENT_RX_CHANNEL_HOPPING_COMPLETE), then fit the TX preamble length to that time (actually you should calculate it for 27 channels, for worst case occasion: if the receiver starts channel scanning on the right channel just before the TX sends the preamble, it will miss the detection, but it should detect after the other 25 channel will have been scanned).

You can monitor RX (and TX) state as well on PRS channel (see how the RAILTEST example configures that, or the Debugging tutorial for reference) to find out how much time the channel switching takes. Those switching times wont change over time, if the channel hopping table is fix (if you are using the same RAIL_RxChannelHoppingConfig_t config). 

For a quick check you can increase the preamble length, let's say, twice as long as you have now, and see whether the number of the missed packets will decrease.

Best Regards,
Bence

Hi Davide,

Which radio events are enabled in your application? Unhandled buffer related interrupts may cause serious issues like getting into default_handler. I suggest to enable all RAIL_EVENTs, and see whether it changes.

The NVIC registers are documented in the Cortex™ -M4 Devices Generic User Guide, and you can find more about how the interrupts are operating in an EFR32MG12 in the reference manual (see section 3.3.1).

Best Regards,
Bence 

Andrew,

If you mean the definition of the LDC_PERIOD in the API docs (the period while the radio is in RX), it is 244 us (8/32768 s) theoretically, but in practice I measured 150 us RX active state on the GPIO. The receiver must receive 20 bits, which takes ~347 us with 57.6 kbps datarate, but if the receiver starts receiving at the middle of the packet (after the preamble sequence), it will miss that packet, therefore the radio must remain longer in RX state, than the minimal LDC period (If we defined it as the minimal time to be able getting preamble detection).

As a result if you are interested in the minimal RX period which is optimal for you usecase, you must first calculate how much energy should be saved both on the TX and on the RX side. This is the trade-off of the LDC applications: one of the two side will consume more power to decrease energy consumption on the other.

Best Regards,
Bence

Content

• Introduction and motivation
• Theory
• Implementation
• Further references

Introduction and motivation

Channel scanning is a receiver side feature. Scanning usually means that a receiver switches its center frequency regularly to be able receiving packets in more frequency bands.

Most of the protocols Silicon Labs RF Software stacks support (BLE, Zigbee, Wi-Fi, Connect) use channel scanning, therefore it is important to get to know more about this topic. The usefulness of this feature is shown through the wide variety of applications, but in this article we will focus only on one type presenting the design steps and showing up several implementation tradeoffs.

First it need to be cleared, what channel does mean in context of Channel scanning. Channel is a medium between the transmitter and receiver side, characterized with physical layer parameters (carrier frequency, modulation type, bitrate, etc.). Mostly two channel differ in center frequency in a concrete protocol. This article will focus on these cases.

We should say some words about the transmitter's side. There may be more transmitter in the same network, all working on its own frequency channel or one which switches between channels (call it as Channel hopping) reducing the chance of interference with other transmitters. Channel hopping implies that one particular channel will be used less than if every packet would be sent on the same channel. Hopping may coexist with LBT and CSMA features also, further decreasing the chance of interference.

Channel hopping and so the scanning can be done either in synchronous way - as FHSS (Frequency Hopping Spread Spectrum) does - or in asynchronous way. Here the synchrony means prior experience at the receiver side. The topic of this article is the asynchronous scanning, when we only known what frequency bands are used by the transmitter(s), but nothing about the time frames, when the channel will be occupied. Our example will show how a Channel scanner can be designed to get every packet.

There is a strong relation between LDC and asynchronous Channel scanning, particularly in the sense of realization. Both has the main concept of detecting empty channel as fast as possible at the receiver side. The purpose of LDC is energy saving in a sparse network, while the purpose of the Channel scanning is rather the better frequency allocation in a dense network, so channel scanning infer a more active receiver with no sleep state.

Design steps for asynchronous Channel scanning

The goal of this section is to show how to build an asynchronous channel scanning application from the beginning assuming ideal transmitter side. We should first define (or get from the standard) the channels with their center frequency. Take care of the false detection on the image frequencies.

In a Channel scanning application the task may not only be to detect the presence of a transmitter but the detection of the lack of any packet in the channel, aiding to switch channel more quickly, raising the chance to capturing a packet (and reducing both sides' power consumption). If it is desired to receive every packet on every channel, some timing consideration should be taken into account.

Since the receiver don't have prior experience about the transmitter, the application should be created with a presumption that every channel will be used with the same chance. In general the receiver scans over a hop table, which tells in what order are the channels scanned.

Receiver's timing considerations

There are several ways to trigger a channel-switching at the receiver side. These triggers can be used together as backup mechanisms as well, to avoid false detection. This is an incomplete list of the possible trigger events:

1. RSSI timeout:
   • can be useful, when packet loss is permissible, and the goal is to detect only the presence of the transmitter,
   • cannot distinguish between desired and undesired signals.
2. Preamble detection timeout:
   • may trigger on invalid preamble pattern as well,
   • the length of the timeout period may depend on AFC (Automatic Frequency Control).
3. Sync word invalid timeout:
   • triggers when invalid or no sync word detected after the preamble,
   • typical use case, if a protocol define similar preamble but different syncwords,
   • it is the longest but most secure option to detect a desired signal.
4. Timing detection timeout:
   • Very similar to Preamble detection timeout,
   • typically it detects absence faster than preamble detection, but causes sensitivity degradation.
5. Digital signal arrival (DSA) detection:
   • usually used for fast absence detection,
   • the fastest, so the most energy efficient option.

Not all features described above are available for every Silicon Labs radios, see the relevant documentation for the possibilities!

In this article we target to use Preamble detection timeout as trigger event. Let's suppose that the transmitter is optimal. This means it transmits long enough packets (most part of the packet is the preamble) on every channel to ensure reception. To quantify this timing requirement the following variables need to introduce:

• T_ri: the receiver's transition time from channel i to the next, (from the end of the RX state to the start of the following one)
• T_detecti: preamble detection timeout (the time of the clear channel detection in this particular case) for the receiver on the i'th channel,
• N: number of channels.

T_ri and T_detecti can be determined empirically and theoretically, respectively.

The former may differ depending on what are the differences between the two channels and how the channel switching functionality is implemented. Therefore it should be measured after defining the hop table.

The latter is described more depth in the LDC KBA at the preamble detection threshold section. The less the configured detection time, the worse the sensitivity, and the accuracy of the detection, but the lower the necessary transmit time as well.

The turnaround time (T_ta), which shows how often will the receiver listen in the same channel, can be calculated by summing T_ri and T_detecti values for all channels.

Transmitter's configuration

In a typical channel scanning application the transmitter transmits relatively long preambles. To ensure detection at the receiver side, the preamble should be on the air at least for one turnaround time. This is true for every channel.

Knowing the data rate at channel i (DR_i, [bit/s]) the preamble length (L_preami, [bit]) can be calculated as follows:

L_preami = T_ta * DR_i.

Summary of design steps

1. Select the number of channels (N) and place them:
   • with the number of channels the RX turnaround time scales linearly,
   • take care of image frequencies.
2. Determine hopping and detection times,
   • from measurements and datasheets.
3. Calculate the minimal transmitted preamble length (in bits) to ensure detection for each channel:
   • the minimal time is the receiver's turnaround time.

Implementation

The channel transition time at the receiver may vary with the implementation: how much time the radio module needs to get the setting parameters (from the host) for channel switching and set those up.

EZRadioPRO

EZRadioPRO supports channel scanning without host MCU control. The RX_HOP property group has a 64 entry RX hop table for predefining channels. Each of the entries hold a channel number corresponding to the channel offset (channel number * channel spacing) from the base frequency, so the channels are placed next to each other by distance of channel spacing. Enabling this "automatic RX Hop" feature the host MCU can switch to sleep state, while the receiver scans all those channels and wakes-up the MCU only when an incoming packet detected (or received).

Other kind of implementation called as "Manual frequency Hopping" is also available on EZRadioPRO devices. The RX_HOP command sets a new listening frequency (usually issued in RX state). With this command channel hop is faster but it needs permanently awake host MCU for scanning.

There are two examples ("Auto frequency hop RX" and "Manual frequency hop RX") among the WDS example projects implementing Auto and Manual hopping applications. We recommend using WDS to generate radio configuration for a manual as well as an automatic RX Channel scanner application, due to the property setting's complexity.

EFR32

The current hopping application example (RAIL Simple TRX MultiPHY) implements frequency hopping with RAIL_Timer and RAIL_StartRx(). It had been developed exactly the same manner as we described above, but its channel configuration has a subGHz and a 2.4 GHz channel with different data rates.

Flex SDK supports also Automatic RX-side channel hopping with APIs RAIL_ConfigRxChannelHopping()and RAIL_EnableRxChannelHopping(). These API's cannot be called when the radio is on. We have an article on how to use those APIs and how to configure the RAIL_RxChannelHoppingConfig_t structure. Note that RX Channel Hopping APIs work on EFR32xG12 and newer parts, but not supported with multiprotocol.

The RAIL API supports to define unlimited number of channels (limited only by the available RAM), and one channel can be defined multiple times, even with different scan times. What's more, RAIL supports more flexible channel definition, so not only the frequency but other Physical layer settings can be defined per channel. After configuring hopping parameters calling RAIL_StartRx() starts on the first channel. There is no need to call other API, the radio automatically hops over the predefined channels. If a packet is received, the radio returns to the idle state (or to the state defined in the auto state transition configuration).

Further References:

• Article series of EZRadioPRO channel scanning feature,
  • Guidance and detailed example to avoid sensing on IF frequency
  • Preamble detection threshold values
  • Trigger scenarios
• RAIL API docs RX Channel Hopping section
• Si446x Datasheet (subsection 5.3.1.1)
• AN633 Programming guide for EZRadioPRO devices (section 10.13 and 10.14)
• RAIL Channel hopping KBA
• Image frequency consideration

Moved to 32-bit MCU Forum.

Andrew,

Sorry, I missed your question last week, your assumption there was correct, as per the datasheet the radio needs 440 us from waking up to get to RX state, you have to take it into account.

Regarding your next post:
- The RX state should be remain longer than the preamble detection threshold (Tron > Tpdp is recommended).
- Don't set preamble detection threshold below 20 bit: it will cause PER issues as the RX couldn't synchronize its clock (couldn't lock reliably) [see datasheet].

And finally my answers to your last questions:
It is always highly recommended to test and confirm the timings and - as you did - the most simple way is to capture those with a logic analyzer. LDC is typically applied with much longer LDC periods therefore the possible configuration has a very coarse resolution at the "low-end" of the potential WUT configurations.

Please increase RX time, while the VALID_PREAMBLE goes high to see whether the LDC is configured incorrectly or we have to find the solution elsewhere.

Best Regards,
Bence

Content

• Motivation
• The Theory
• Realizing in RAIL
• Examples
  • Typical use cases
  • Further development

Motivation

I used to play quiz game with my friends, but when we needed to determine who was the first, who put his hand after a question arose, we get into disagreement. For this I was long thinking about a simple wireless system, what is able to decide who will be the responder, taking the responsibility (and our annoyance as well) out. This system should consist of one central device and any number of nodes, enhancing the network scalability, and the synchronization process should be simple and quick as well (capable to be done even for every new question), enabling simultaneous synchronization. This task may be done through existing solutions like Bluetooth or Wi-Fi, but I want accuracy in the order of microseconds, to be as precise as possible. RAIL time base is suitable for this, as it is in microseconds. It's not the best example to show the strength of synchronization, but I will cover here in this KBA some other examples.

This KBA presents a solution to synchronize networks containing one Master device and one or more Slave devices.

Application examples

There are many examples around us to everyday use of time synchronization including but not limited to the following.

GPS and other localization examples

Probably the most known application example of time synchronization is GPS (Global Positioning System). The synchronization is necessary to measure distance with wireless devices, as the distance is derived from the amount of time, since we know the velocity of light spreading through the air (vacuum). And if we can measure distance the next step is the positioning or localization.

Multihop network synchronization

If a computer connects to the internet, there is an opportunity right away to synchronize it to the global time (NTP). Since it is a multihop network, the delay from the reference time does increase, so the accuracy decrease with the number of hops.

Similarly, if the task is the scheduling a bunch of sensor nodes spanning a large geographical area it can be solved with time synchronization, where the nodes synchronize their neighbor nodes.

Low duty cycle

If you are not familiar with the term of Low duty cycle (LDC) take a look at this article.

This technology helps to save energy using lower power energy modes with a less accurate timer. There may be need for regular synchronization between the nodes, to minimize the active time along with the consumption, though this process handled automatically due to the communication.

The Theory of Operation

In the context of time synchronization it is good enough for the most application to have one reference time, which is known to all device, and in most cases this is the time of one particular device which dictates the system time. It gives the opportunity to schedule events in a common time base. We will refer to this device, which serves the reference time, as Master, and to the others as Slaves.

The synchronization means that one Slave device tries to estimate the Master's time, and then its operation scheduled in that assumed time base. We intentionally used the 'try' verb, because a synchronization is always inaccurate and requires regular adjustments. Though a simple synchronization can be done very quickly, the hardest part of the process is the evaluation of this inaccuracy, and the (sequential/continuous) correction, especially in a dense network, with a couple of Slave devices.

Time synchronization involves a several new terms which are describing delay times. Without completeness those are:

1. Warmup time (including channel access time): the time from assembling packet, through requesting the radio module to send, to the start of the first bit's transmission.
2. Transmission time: the time while the packet is on-the-air, from the first bit (end of Warmup time) to the last.
3. Propagation time: the time starting from related parts of the packet leave the transmitter and reach receiver (relative: doesn't have beginning nor end).
4. Reception time: similarly to the Transmission time, the time while the packet is on-the-air, till the end of the last bit of packet have been totally received.
5. Processing time: time to process received packet.
6. Settling time: time to set the time base at the receiver to the assumed time base of the transmitter (have not found this time in the literature, but the example application use this).

Using these delay terms and the terms describing difference of two clocking system the inaccuracy can be quantized.

Errors in/during/after synchronization

In terms of Time synchronization two types of error could occur. One is the offset error and the other is the dynamic or slope error (clock drift). Synchronization is nothing more than the elimination of these errors.

The offset error is very straightforward, it is the error measured between two devices right after the synchronization. This kind of error come from the incorrect calculation of delay times. It is a zero-order (constant) error.

The dynamic error is the speed difference between the clocks of the two devices. This is caused by the speed difference of the two oscillator. The dynamic error can be reduced by calibrating the clocking systems to a common reference (e.g. CTune calibration). The quality of this error is in first-order.

These errors might change during the lifetime of a system: as the temperature changes, the oscillators' frequencies change as well. This may be compensable with CTune calibration.

Measuring the errors

The task of synchronizing is very similar to a geometry exercise, finding a straight line in a coordinate system. This task requires a point and a slope or two points in a defined coordinate system. In this metaphor the coordinate system is defined by the "real time" and the Master's time base (assuming the Master device uses a 0ppm clock).

In the figures below the x axis represents the real time, and the y is the (discrete) value of time with the devices' metric (RAIL Time base, which is actually 1μs), the solid line represents the Master's time while the dotted line is the Slave's time.

Offset error compensation

Let's assume first that the slopes are the same (the clocks run at the same speed), therefore we need only a well defined point for the line fitting. With this assumption the offset error can be compensated in the following way:

1. Generate an event perceived by the master and slave devices
2. Get the event's timestamps on both side (~get to know the coordinate system)
3. Share Master's timestamp with the Slave(s) (~get to know the point on the coordinate system)
4. Compensate offset error on Slave(s)' side (~place the line on the coordinate system)

The task is to be able to figure out the what is the time on both the Master and the Slave at a given time, using the Slave's time base - It will be different, since they weren't powered up at the same time, thus the counters hold different values.

As a first step it should happen such an event which is perceived on both the Master and the Slave device. It is evident that this event may be a packet transmission and reception, therefore the delay times have to be known. As a result two timestamp generated on each device. Such an event will provide enough information to get the offset error. The clearest way to send a message from the Master, which is received by the Slave(s).

The next step is to share the Master's timestamp with the Slave device(s). It means that Master should send an another message, which holds the previous message's timestamp. From this message, the Slave(s) will know much its local clock differs from the Master's clock - the offset error - and will know more or less what time is it on the other at present (with the initial assumption).

Dynamic error compensation

In the real world there is always some difference between the clock systems. This is where the dynamic error takes place, and we should compensate it as well.

This can be done in a very similar way as we saw it in the previous section: we should repeat the process. This means getting an another point on the coordinate system, and from this information the to get the first-order error (the slope, or we can say that the velocity of the offset error's increase) can be defined and the dynamic error can be eliminated.

This process requires only 3 packet to be sent from a Master device, as the second packet, which holds the first packet's timestamp has it's own timestamp, which can be sent in the third packet.

Once these errors are known, they can be eliminated by calculation (software) or by modifying the clocking system (hardware) to be synchronized.

Implementation in RAIL

Possibilities

RAIL supports timestamping for both the transmitted and the received packet. The enum RAIL_PacketTimePosition_t defines the possibilities timestamp positions. The available positions are:

• the start of the preamble (first preamble bit sent or received)
• end of the syncword (last syncword bit sent or received),
• end of the packet (last bit of the packet sent or received).

In fact, at TX side only the end of the packet will be recorded as timestamp, and the other timestamps calculated relatively from this value. This implies additive error due to dynamic error in case of a very long packet (though it is almost negligible). Furthermore this means that there is no way to timestamp the packet, then send out the timestamp in that same packet.

NOTE: This will change in the future on some devices!

Since we won't use long packets in such an application, and this difference from dynamic error would be eliminated with the offset error compensation, we chose the timestamp position to the end of syncword (though it does not have any benefit).

The synchronization process

This image shows the synchronization process:

The red part is the preamble and the syncword of the packet. The Slave must save the timestamp of first packet's end (T_s) as well as the Master (T_m). The difference between the two timestamp is the propagation time (T_prop). Then in the second packet the Master device should send T_m, which is in its own time base. Finally the Slave device receiving T_m will know the difference between its own and the Master's time base eliminating the offset error, whic is the first step of synchronization.

But then, after a while, the error between the time assumed by the Slave, and the real time according to the Master's time base will differ due to HFXO clock deviation. This can be reduced repeating the previous step and "predicting" the masters new T_m from the last known value. The difference of the actual and the predicted T_m shows whether the Slave clocks are slower or the Master's. That way the slope error can be estimated. The accuracy of the estimation improves with the increasing delay between the synchronizations.

One can add the question: what about the delay times? We did not care about most of them (ones have definite beginning and end), due to the timestamp system of RAIL eliminates them. We didn't care about the propagation time either, since it is less than 1 us (what is the resolution of our timer) if the distance between the two device is less than 300 m.

Explanation of the code

The application is based on the example "simple-rail-with-hal" for BRD4169B and BRD4164A radio boards (EFR32MG14/MG12) which was used with WSTK. The application is using the Command interpreter Plugin, thus "Virtual COM Port" has to be enabled in the hardware configurator. No other modification should be made in the radio and hardware configurator.

The application was extended with two additional packet: One from the Slave with an assumed value (T_predict) to the Master's second packet's timestamp, and with a compensation feedback (T_comp) from the Master, which tells the Salve how accurate the prediction was.

Caution: This is only a demonstration project, we want to show the synchronization process in the simplest way, so our implementation can be improved. Furthermore it contains features assisting debug. Here are some notes what you should avoid in your design.

• Don't use the RAIL_SetTime() API in that way, as it is used here.
• The blocking delay implementation is not recommended. A timer interrupt routine may be more elegant in your project.
• Starting Master without Slave (or vice versa) will stuck while it won't get any packet. You should reset the device to exit this state.

States and roles

The application operates upon a state machine, equipped with command line interface, GPIO interrupt handlers for push buttons, and with radio event handling functionality. 5 state was implemented for both Master and Slave(s). Those are matching to each other. These are:

1. Waiting for start (pushbutton or CLI command),
2. Sending/receiving first packet from Master (produce timestamps T_m_1 and T_s_1),
3. Sending/receiving second packet from Master (produce T_m_2 and T_s_2 and send/get T_m_1),
4. Sending/receiving first packet from Slave (send T_predict for predicting T_m_2),
5. Sending/receiving third packet from Master (get T_comp).

A single application has been developed with two roles: there is a run-time variable to select whether a device is Master (default) or Slave.

typedef enum MasterStates {
   MASTER_WAIT, MASTER_STATE_SEND_FIRST_PACKET, MASTER_STATE_SEND_SECOND_PACKET,
} MasterStates_t;

typedef enum MasterStates {
  MASTER_WAIT, MASTER_STATE_SEND_FIRST_PACKET, MASTER_STATE_SEND_SECOND_PACKET,
} MasterStates_t;

typedef enum Role {
  ROLE_MASTER, ROLE_SLAVE
} Role_t;
...
static Role_t role = ROLE_MASTER;

The CLI commands

The application implements 5 cli commands:

• cliSetRole selects between Master (0) and Slave (1) roles,
• cliStart starts a synchronization process (if the argument is 1),
• cliSetCtune sets the CTune value (second argument); if the first argument is 0 the second argument will be used as CTune, if 1 the second argument will be saved to the User Page first, if 2 the CTune value will be fetched from the User Page and the second argument will not take any effect,
• cliGetTime prints the current time,
• cliGetVal prints one of the timestamps or calculations (for debug use only).

The interrupt handler routines

The interrupt handling routines are mostly setting flags which are monitored in the main loop.

static volatile bool sent = false;
static volatile bool received = false;
static volatile bool start = false;

The left pushbutton (PB1) does exactly the same thing as cliStart command does, while the right (PB0) operates as cliGetTime.

void gpioCallback(uint8_t pin) {
  if (pin & 0x01) {
    start = true;
  } else {
    RAIL_Time_t t = RAIL_GetTime();
    printf("time %d\n", t);
  }
}

The enabled radio events are the RX and TX related ones. Only for the successful events are handled in this application, RAIL_EVENT_TX_PACKET_SENT and RAIL_EVENT_RX_PACKET_RECEIVED. Every packet transmission and reception refreshes the timestamp value bound to the syncword's last bit.

static volatile RAIL_PacketTimeStamp_t timeStamp;
...

RAIL_Events_t events = (RAIL_EVENTS_TX_COMPLETION | RAIL_EVENTS_RX_COMPLETION);
RAIL_ConfigEvents(railHandle, events, events);

...

void RAILCb_Generic(RAIL_Handle_t railHandle, RAIL_Events_t events) {
  
  if (events & RAIL_EVENT_TX_PACKET_SENT) {
    // Set flag for the main loop
    sent = true;

    RAIL_TxPacketDetails_t packetDetails = { .timeSent = timeStamp .isAck = false };
    RAIL_GetTxPacketDetails(railHandle, &packetDetails);

    timeStamp = packetDetails.timeSent;
    // Don't print in the event handler! We do that for demonstration only.
    printf("packet sent\n");
  }

  if (events & RAIL_EVENT_RX_PACKET_RECEIVED) {
    // Set flag for the main loop
    received = true;

    // Get the newest packet's information from RX FIFO
    RAIL_RxPacketInfo_t packetInfo;
    RAIL_RxPacketHandle_t packetHandle = RAIL_RX_PACKET_HANDLE_NEWEST;
    RAIL_GetRxPacketInfo(railHandle, packetHandle, &packetInfo);

    // Set the position of the packet where the timestamp should be acquired,
    // and the packet's length to calculate it correctly
    timeStamp.timePosition = RAIL_PACKET_TIME_AT_PACKET_END;
    timeStamp.totalPacketBytes = packetInfo.packetBytes;
    RAIL_RxPacketDetails_t packetDetails = { .timeReceived = timeStamp };
    RAIL_GetRxPacketDetails(railHandle, packetHandle, &packetDetails);

    RAIL_CopyRxPacket(rxPacket, &packetInfo);

    RAIL_ReleaseRxPacket(railHandle, packetHandle);

    timeStamp = packetDetails.timeReceived;
    RAIL_ResetFifo(railHandle, false, true);
    // Don't print in the event handler! We do that for demonstration only.
    printf("packet received %d\n", timeStamp.packetTime);
  }

}

Manual calibration

In this section we will present how the application works.

After flashing the binary to two device we connected their serial ports to a PC. In the serial terminal we should only issue a setRole 1 command to select the slave device. Push the PB1 first on the slave (this will enable RX) and then on the master (send the first message). After some milliseconds the two device will be synchronized.

We can prove that the the devices are synchronized by getting the current time from both device in the same time. This can be done with pushing PB0, while the two GPIO lines are connected. The button's line can be found at the seventh pin of the external pin header connector on the WSTK (using the BRD4169B radio boards).

I pushed PB0 per ~1 second, and the two terminal windows printed the following timestamps (in microseconds):

The first time pair shows that the slave counts 3 more us than the master device, and in ~29 seconds the difference increased to 16 us. It means that the slave's clock ran faster causing 1 us drift per ~2.25 second (slope error).

Adjusting CTune (with the SetCtune command) this kind of error can be minimized (increasing on the slave to slow down, or decreasing on the master to speed up the oscillator).

This accuracy is now good enough for my quiz application.

Content

• Motivation
• The Theory
• Realizing in RAIL
• Examples
  • Typical use cases
  • Further development

Motivation

I used to play quiz game with my friends, but when we needed to determine who was the first, who put his hand after a question arose, we get into disagreement. For this I was long thinking about a simple wireless system, what is able to decide who will be the responder, taking the responsibility (and our annoyance as well) out. This system should consist of one central device and any number of nodes, enhancing the network scalability, and the synchronization process should be simple and quick as well (capable to be done even for every new question), enabling simultaneous synchronization. This task may be done through existing solutions like Bluetooth or Wi-Fi, but I want accuracy in the order of microseconds, to be as precise as possible. RAIL time base is suitable for this, as it is in microseconds. It's not the best example to show the strength of synchronization, but I will cover here in this KBA some other examples.

This KBA presents a solution to synchronize networks containing one Master device and one or more Slave devices.

Application examples

There are many examples around us to everyday use of time synchronization including but not limited to the following.

GPS and other localization examples

Probably the most known application example of time synchronization is GPS (Global Positioning System). The synchronization is necessary to measure distance with wireless devices, as the distance is derived from the amount of time, since we know the velocity of light spreading through the air (vacuum). And if we can measure distance the next step is the positioning or localization.

Multihop network synchronization

If a computer connects to the internet, there is an opportunity right away to synchronize it to the global time (NTP). Since it is a multihop network, the delay from the reference time does increase, so the accuracy decrease with the number of hops.

Similarly, if the task is the scheduling a bunch of sensor nodes spanning a large geographical area it can be solved with time synchronization, where the nodes synchronize their neighbor nodes.

Low duty cycle

If you are not familiar with the term of Low duty cycle (LDC) take a look at this article.

This technology helps to save energy using lower power energy modes with a less accurate timer. There may be need for regular synchronization between the nodes, to minimize the active time along with the consumption, though this process handled automatically due to the communication.

The Theory of Operation

In the context of time synchronization it is good enough for the most application to have one reference time, which is known to all device, and in most cases this is the time of one particular device which dictates the system time. It gives the opportunity to schedule events in a common time base. We will refer to this device, which serves the reference time, as Master, and to the others as Slaves.

The synchronization means that one Slave device tries to estimate the Master's time, and then its operation scheduled in that assumed time base. We intentionally used the 'try' verb, because a synchronization is always inaccurate and requires regular adjustments. Though a simple synchronization can be done very quickly, the hardest part of the process is the evaluation of this inaccuracy, and the (sequential/continuous) correction, especially in a dense network, with a couple of Slave devices.

Time synchronization involves a several new terms which are describing delay times. Without completeness those are:

1. Warmup time (including channel access time): the time from assembling packet, through requesting the radio module to send, to the start of the first bit's transmission.
2. Transmission time: the time while the packet is on-the-air, from the first bit (end of Warmup time) to the last.
3. Propagation time: the time starting from related parts of the packet leave the transmitter and reach receiver (relative: doesn't have beginning nor end).
4. Reception time: similarly to the Transmission time, the time while the packet is on-the-air, till the end of the last bit of packet have been totally received.
5. Processing time: time to process received packet.
6. Settling time: time to set the time base at the receiver to the assumed time base of the transmitter (have not found this time in the literature, but the example application use this).

Using these delay terms and the terms describing difference of two clocking system the inaccuracy can be quantized.

Errors in/during/after synchronization

In terms of Time synchronization two types of error could occur. One is the offset error and the other is the dynamic or slope error (clock drift). Synchronization is nothing more than the elimination of these errors.

The offset error is very straightforward, it is the error measured between two devices right after the synchronization. This kind of error come from the incorrect calculation of delay times. It is a zero-order (constant) error.

The dynamic error is the speed difference between the clocks of the two devices. This is caused by the speed difference of the two oscillator. The dynamic error can be reduced by calibrating the clocking systems to a common reference (e.g. CTune calibration). The quality of this error is in first-order.

These errors might change during the lifetime of a system: as the temperature changes, the oscillators' frequencies change as well. This may be compensable with CTune calibration.

Measuring the errors

The task of synchronizing is very similar to a geometry exercise, finding a straight line in a coordinate system. This task requires a point and a slope or two points in a defined coordinate system. In this metaphor the coordinate system is defined by the "real time" and the Master's time base (assuming the Master device uses a 0ppm clock).

In the figures below the x axis represents the real time, and the y is the (discrete) value of time with the devices' metric (RAIL Time base, which is actually 1μs), the solid line represents the Master's time while the dotted line is the Slave's time.

Offset error compensation

Let's assume first that the slopes are the same (the clocks run at the same speed), therefore we need only a well defined point for the line fitting. With this assumption the offset error can be compensated in the following way:

1. Generate an event perceived by the master and slave devices
2. Get the event's timestamps on both side (~get to know the coordinate system)
3. Share Master's timestamp with the Slave(s) (~get to know the point on the coordinate system)
4. Compensate offset error on Slave(s)' side (~place the line on the coordinate system)

The task is to be able to figure out the what is the time on both the Master and the Slave at a given time, using the Slave's time base - It will be different, since they weren't powered up at the same time, thus the counters hold different values.

As a first step it should happen such an event which is perceived on both the Master and the Slave device. It is evident that this event may be a packet transmission and reception, therefore the delay times have to be known. As a result two timestamp generated on each device. Such an event will provide enough information to get the offset error. The clearest way to send a message from the Master, which is received by the Slave(s).

The next step is to share the Master's timestamp with the Slave device(s). It means that Master should send an another message, which holds the previous message's timestamp. From this message, the Slave(s) will know much its local clock differs from the Master's clock - the offset error - and will know more or less what time is it on the other at present (with the initial assumption).

Dynamic error compensation

In the real world there is always some difference between the clock systems. This is where the dynamic error takes place, and we should compensate it as well.

This can be done in a very similar way as we saw it in the previous section: we should repeat the process. This means getting an another point on the coordinate system, and from this information the to get the first-order error (the slope, or we can say that the velocity of the offset error's increase) can be defined and the dynamic error can be eliminated.

This process requires only 3 packet to be sent from a Master device, as the second packet, which holds the first packet's timestamp has it's own timestamp, which can be sent in the third packet.

Once these errors are known, they can be eliminated by calculation (software) or by modifying the clocking system (hardware) to be synchronized.

Implementation in RAIL

Possibilities

RAIL supports timestamping for both the transmitted and the received packet. The enum RAIL_PacketTimePosition_t defines the possibilities timestamp positions. The available positions are:

• the start of the preamble (first preamble bit sent or received)
• end of the syncword (last syncword bit sent or received),
• end of the packet (last bit of the packet sent or received).

In fact, at TX side only the end of the packet will be recorded as timestamp, and the other timestamps calculated relatively from this value. This implies additive error due to dynamic error in case of a very long packet (though it is almost negligible). Furthermore this means that there is no way to timestamp the packet, then send out the timestamp in that same packet.

NOTE: This will change in the future on some devices!

Since we won't use long packets in such an application, and this difference from dynamic error would be eliminated with the offset error compensation, we chose the timestamp position to the end of syncword (though it does not have any benefit).

The synchronization process

This image shows the synchronization process:

The red part is the preamble and the syncword of the packet. The Slave must save the timestamp of first packet's end (T_s) as well as the Master (T_m). The difference between the two timestamp is the propagation time (T_prop). Then in the second packet the Master device should send T_m, which is in its own time base. Finally the Slave device receiving T_m will know the difference between its own and the Master's time base eliminating the offset error, whic is the first step of synchronization.

But then, after a while, the error between the time assumed by the Slave, and the real time according to the Master's time base will differ due to HFXO clock deviation. This can be reduced repeating the previous step and "predicting" the masters new T_m from the last known value. The difference of the actual and the predicted T_m shows whether the Slave clocks are slower or the Master's. That way the slope error can be estimated. The accuracy of the estimation improves with the increasing delay between the synchronizations.

One can add the question: what about the delay times? We did not care about most of them (ones have definite beginning and end), due to the timestamp system of RAIL eliminates them. We didn't care about the propagation time either, since it is less than 1 us (what is the resolution of our timer) if the distance between the two device is less than 300 m.

Explanation of the code

The application is based on the example "simple-rail-with-hal" for BRD4169B and BRD4164A radio boards (EFR32MG14/MG12) which was used with WSTK. The application is using the Command interpreter Plugin, thus "Virtual COM Port" has to be enabled in the hardware configurator. No other modification should be made in the radio and hardware configurator.

The application was extended with two additional packet: One from the Slave with an assumed value (T_predict) to the Master's second packet's timestamp, and with a compensation feedback (T_comp) from the Master, which tells the Salve how accurate the prediction was.

Caution: This is only a demonstration project, we want to show the synchronization process in the simplest way, so our implementation can be improved. Furthermore it contains features assisting debug. Here are some notes what you should avoid in your design.

• Don't use the RAIL_SetTime() API in that way, as it is used here.
• The blocking delay implementation is not recommended. A timer interrupt routine may be more elegant in your project.
• Starting Master without Slave (or vice versa) will stuck while it won't get any packet. You should reset the device to exit this state.

States and roles

The application operates upon a state machine, equipped with command line interface, GPIO interrupt handlers for push buttons, and with radio event handling functionality. 5 state was implemented for both Master and Slave(s). Those are matching to each other. These are:

1. Waiting for start (pushbutton or CLI command),
2. Sending/receiving first packet from Master (produce timestamps T_m_1 and T_s_1),
3. Sending/receiving second packet from Master (produce T_m_2 and T_s_2 and send/get T_m_1),
4. Sending/receiving first packet from Slave (send T_predict for predicting T_m_2),
5. Sending/receiving third packet from Master (get T_comp).

A single application has been developed with two roles: there is a run-time variable to select whether a device is Master (default) or Slave.

typedef enum MasterStates {
   MASTER_WAIT, MASTER_STATE_SEND_FIRST_PACKET, MASTER_STATE_SEND_SECOND_PACKET,
} MasterStates_t;

typedef enum MasterStates {
  MASTER_WAIT, MASTER_STATE_SEND_FIRST_PACKET, MASTER_STATE_SEND_SECOND_PACKET,
} MasterStates_t;

typedef enum Role {
  ROLE_MASTER, ROLE_SLAVE
} Role_t;
...
static Role_t role = ROLE_MASTER;

The CLI commands

The application implements 5 cli commands:

• cliSetRole selects between Master (0) and Slave (1) roles,
• cliStart starts a synchronization process (if the argument is 1),
• cliSetCtune sets the CTune value (second argument); if the first argument is 0 the second argument will be used as CTune, if 1 the second argument will be saved to the User Page first, if 2 the CTune value will be fetched from the User Page and the second argument will not take any effect,
• cliGetTime prints the current time,
• cliGetVal prints one of the timestamps or calculations (for debug use only).

The interrupt handler routines

The interrupt handling routines are mostly setting flags which are monitored in the main loop.

static volatile bool sent = false;
static volatile bool received = false;
static volatile bool start = false;

The left pushbutton (PB1) does exactly the same thing as cliStart command does, while the right (PB0) operates as cliGetTime.

void gpioCallback(uint8_t pin) {
  if (pin & 0x01) {
    start = true;
  } else {
    RAIL_Time_t t = RAIL_GetTime();
    printf("time %d\n", t);
  }
}

The enabled radio events are the RX and TX related ones. Only for the successful events are handled in this application, RAIL_EVENT_TX_PACKET_SENT and RAIL_EVENT_RX_PACKET_RECEIVED. Every packet transmission and reception refreshes the timestamp value bound to the syncword's last bit.

static volatile RAIL_PacketTimeStamp_t timeStamp;
...

RAIL_Events_t events = (RAIL_EVENTS_TX_COMPLETION | RAIL_EVENTS_RX_COMPLETION);
RAIL_ConfigEvents(railHandle, events, events);

...

void RAILCb_Generic(RAIL_Handle_t railHandle, RAIL_Events_t events) {
  
  if (events & RAIL_EVENT_TX_PACKET_SENT) {
    // Set flag for the main loop
    sent = true;

    RAIL_TxPacketDetails_t packetDetails = { .timeSent = timeStamp .isAck = false };
    RAIL_GetTxPacketDetails(railHandle, &packetDetails);

    timeStamp = packetDetails.timeSent;
    // Don't print in the event handler! We do that for demonstration only.
    printf("packet sent\n");
  }

  if (events & RAIL_EVENT_RX_PACKET_RECEIVED) {
    // Set flag for the main loop
    received = true;

    // Get the newest packet's information from RX FIFO
    RAIL_RxPacketInfo_t packetInfo;
    RAIL_RxPacketHandle_t packetHandle = RAIL_RX_PACKET_HANDLE_NEWEST;
    RAIL_GetRxPacketInfo(railHandle, packetHandle, &packetInfo);

    // Set the position of the packet where the timestamp should be acquired,
    // and the packet's length to calculate it correctly
    timeStamp.timePosition = RAIL_PACKET_TIME_AT_PACKET_END;
    timeStamp.totalPacketBytes = packetInfo.packetBytes;
    RAIL_RxPacketDetails_t packetDetails = { .timeReceived = timeStamp };
    RAIL_GetRxPacketDetails(railHandle, packetHandle, &packetDetails);

    RAIL_CopyRxPacket(rxPacket, &packetInfo);

    RAIL_ReleaseRxPacket(railHandle, packetHandle);

    timeStamp = packetDetails.timeReceived;
    RAIL_ResetFifo(railHandle, false, true);
    // Don't print in the event handler! We do that for demonstration only.
    printf("packet received %d\n", timeStamp.packetTime);
  }

}

Manual calibration

In this section we will present how the application works.

After flashing the binary to two device we connected their serial ports to a PC. In the serial terminal we should only issue a setRole 1 command to select the slave device. Push the PB1 first on the slave (this will enable RX) and then on the master (send the first message). After some milliseconds the two device will be synchronized.

We can prove that the the devices are synchronized by getting the current time from both device in the same time. This can be done with pushing PB0, while the two GPIO lines are connected. The button's line can be found at the seventh pin of the external pin header connector on the WSTK (using the BRD4169B radio boards).

I pushed PB0 per ~1 second, and the two terminal windows printed the following timestamps (in microseconds):

The first time pair shows that the slave counts 3 more us than the master device, and in ~29 seconds the difference increased to 16 us. It means that the slave's clock ran faster causing 1 us drift per ~2.25 second (slope error).

Adjusting CTune (with the SetCtune command) this kind of error can be minimized (increasing on the slave to slow down, or decreasing on the master to speed up the oscillator).

This accuracy is now good enough for my quiz application.