 |
USRP Hardware Driver and USRP Manual
Version: 4.4.0.HEAD-0-g5fac246b
UHD and USRP Manual
|
|
|
USRP Hardware Driver and USRP Manual
Version: 4.4.0.HEAD-0-g5fac246b
UHD and USRP Manual
|
|
One of the mechanisms that RFNoC blocks can use to control their configuration are properties. There are two types of properties within a block: User properties, and edge properties.
To illustrate these types of properties, we start with an example: Take DDC block (uhd::rfnoc::ddc_block_control), which can take in a signal of a certain sampling rate, shift it in frequency, and resample to a lower sampling rate.
First, we take a look at the user properties. User properties are properties which are attributes of a block, they simply define certain characteristics of the block. Here, we focus on two user properties: The frequency and the decimation. The former is a floating point value which stores, in Hz, the amount by which the incoming signal is shifted in frequency. The decimation is the integer value by which the incoming sampling rate is divided. Put differently, the user properties control the behaviour of the DDC block.
On top of that, the DDC block also has an edge property for the sampling rate on both the outgoing edge and the incoming edge. This property describes something about the data on this graph edge.
The key thing, which separates properties from other attributes associated with the block, is that we can describe relationships between the properties. In this example, the following relationships exist:
samp_rate property on the outgoing edge equals the samp_rate property on the incoming edge divided by the decim value.freq user property by the samp_rate edge property on the incoming edge.User properties also act as an API for RFNoC blocks. It is possible to access user properties using the uhd::rfnoc::node_t::get_property() and uhd::rfnoc::node_t::set_property() and uhd::rfnoc::node_t::set_properties() APIs. A list of user properties can be read by calling uhd::rfnoc::node_t::get_property_ids().
Often, block controllers will provide explicit C++ APIs on top of the user properties. For example, the DDC block control has an API call uhd::rfnoc::ddc_block_control::set_freq(), which provides additional features on top of the user property (e.g., it can also set a command time). Direct access of user properties is thus often not necessary, and can be considered a lower-level API than directly calling block APIs.
By itself, these relationships between properties are already useful to describe something about the block, but the real power comes when connecting blocks. This allows blocks to communicate their settings, even without knowing which block they're connected to. Consider the following example, using a radio block (uhd::rfnoc::radio_control), the DDC block, and a hypothetical Modem Block, which expects incoming samples at a fixed data rate of 20 Msps. Assume we are running this flow graph in a USRP X310, with a master clock rate of 200 Msps:
Upon initialization, the radio block will configure its hardware clocks, and set a sampling rate of 200 Msps on its outgoing edge property (samp_rate). This property is also the incoming edge property of the DDC block, which means it is now initialized its sampling rate to 200 Msps. Setting the freq property will now work accurately, because it can calculate normalized frequencies.
The Modem Block will set its incoming edge property samp_rate to 20 Msps, which is also the outgoing edge property of the DDC block. The DDC block must now find a way to reconcile these edge properties, which it can achieve by setting the decim user property to 20. Now, the DDC has automatically set a decimation without anyone calling into its rate-changing APIs, and without the Modem Block or the Radio Block knowing anything about how the DDC block works.
Similarly, if the Modem Block would have requested a sampling rate of 30.72 Msps, the DDC would not have been able to resolve the incoming and outgoing sampling rates with an integer decimation factor. It would have requested the radio to produce an integer multiple of 30.72 Msps, which the radio can't provide. Using properties, UHD would be able to throw an exception, telling the user that the requested combination is not achievable.
The DDC block is configurable to have a variable number of channels, and the same is true for the properties. To accommodate for this, both edge and user properties have an index value. A multi-channel DDC block is more accurately depicted as such with regard to its properties:
Note: The index of edge properties is tightly coupled to the actual edge being used (i.e., samp_rate on input edge 0 also has a property index of 0). However, user properties do not enforce this relationship. In the DDC block, it would be possible for the user property decim[0] to control channel 1 (which it doesn't). This is useful for properties that don't map directly to channels (e.g., a simpler version of the DDC block might only have a single decim user property for all channels).
In summary, to accurately identify a property, not only its name is required, but also its type (input edge, user property, or output edge) and its index. For the sake of convenience, the type and index are sometimes combined into a uhd::rfnoc::res_source_info object.
To define properties in an RFNoC block, the block controller needs to declare class attributes of type uhd::rfnoc::property_t. This is a template type which can take in different types of C++ types (e.g., the decim user property is an integer type, and the freq user property is of type double). Upon creation of the class attribute, properties are configured with their property name (e.g. freq, decim), optionally a default value, their type (user or edge property), and their index. The uhd::rfnoc::ddc_block_control defines its frequency property attribute variable as such:
Next, the attribute needs to be initialized, like any other class attribute. If we were to hard-code a 2-channel DDC with a default frequency shift of 0 Hz, this would be a valid initialization:
Note: It is important to choose a container for properties which does not modify their memory locations! If a std::vector<> is chosen as a container, it is required to reserve() memory before filling the container.
Up until now, the property is just a class attribute. For UHD to make it available for property propagation, it needs to be registered by calling uhd::rfnoc::node_t::register_property():
The final step of creating properties is to define the relationship between them, which is covered in the following section.
A property resolver is a function object that is associated with an input list of properties and an output list of properties. The function object will be executed when any property on the input list changes. It may modify any property in the output list. For more details, see uhd::rfnoc::node_t::add_property_resolver().
As an example, here is a shortened version of one of the resolvers in uhd::rfnoc::ddc_block_control:
A few noteworthy comments:
decim, it is still possible to coerce decim to a different, valid value.samp_rate_in, samp_rate_out, and decim are in a consistent state.To keep track of which properties have been modified, every property has a "dirty" flag that is set when a property's value is changed. This dirty flagged is also used to determine which property resolver functions need to be called. A property resolver may modify (and thus dirty) other properties, but resolvers may never set properties to conflicting values. For example, the DDC block controller has a different resolver that is called when its input sampling rate is modified, but it uses the same equations to relate decimation, input rate, and output rate so the resulting values are consistent.
Imagine a block that has both an fft_size and bin_width user property. They are related by bin_width = sampling_rate_in / fft_size, where sampling_rate_in is an input edge property. Adding two separate resolvers like this is valid:
Both resolvers will modify the input for the other resolver. However, because they use the same equations, this does not incur a circular resolution.
Note: When using floating-point type properties, it is possible to incur a circular resolution when floating point rounding errors occur, so they need to be accounted for. See, for example, the implementation of uhd::rfnoc::ddc_block_control.
When all resolvers with dirty inputs have been run, and no conflicts have occurred, properties are flagged clean. When this happens, a clean callback is executed which can trigger further actions (it may not, however, modify properties. See uhd::rfnoc::node_t::register_property()).
As an example, assume a block configures an FFT size. It first checks the FFT size is as power of two, then writes the log2 of the FFT size to a hardware register.
The following two implementations have the same effect. First, we use the resolver to write to the hardware:
Alternatively, we poke the register as part of cleaning the property:
The differences are subtle:
In this particular instance, the second approach is the more readable and is recommended. However, not always does changing properties map to poking registers (or other actions) in a clean, one-to-one manner, which is when the first approach may be better suited.
In an RFNoC graph, after calling uhd::rfnoc::rfnoc_graph::commit(), edge properties are used to resolve properties of a whole graph. When any property is changed, the corresponding block's properties are resolved as explained before. However, by modifying the edge properties of a block, other blocks' properties may be dirtied as well. In this case, the resolver algorithm will keep resolving blocks until either all properties are clean, or a conflict is detected.
The graph object uses a topological sort to identify the order in which blocks are resolved. However, in RFNoC, it is OK to have loops, or back edges.
Consider the following graph:
In this application, we receive a signal, use the DDC block to correct a frequency offset digitally, and then apply some custom DSP before transmitting the signal again, using the same radio block. In order to create a valid graph that UHD can handle, one of the edges needs to be declared a back-edge.
When declaring an edge as a back-edge, this signals to UHD that this edge should not be used for viewing the graph as a directed, acyclic graph (DAG), which is important during property propagation. In the graph above, UHD would sort blocks topologically (Radio Block, DDC Block, Custom DSP Block) by ignoring back-edges. Then, properties are forwarded in topological order first across forward edges, then across back edges.
This can lead to resolution errors. For example, if the DDC were to be configured to decimate the sampling rate, and the custom DSP block would not correct for that, the output edge property samp_rate of the custom DSP block would be mismatched compared to the input edge property samp_rate of the radio. This could either re-trigger a new cascade of property settings, which means the algorithm can't converge, and UHD would detect that and throw an exception. Or the radio block might not accept a sample rate update on its input port, which would also trigger an exception.
It is highly recommended to not declare edges as back-edges unless necessary. In the graph above, any of the three edges could have been declared a back-edge to result in a topologically valid graph that can still resolve its properties, but the one chosen is the most intuitive selection.
Since every block can define its own edge- and user properties, it is likely that a block may not have defined an edge property that an up- or downstream block has.
Consider the example of the previous section. The "Custom DSP" block may not have defined edge properties for the sampling rate. The Radio and DDC blocks however, do have a samp_rate edge property defined.
The way blocks handle such properties is by setting a forwarding policy (see uhd::rfnoc::node_t::set_prop_forwarding_policy() and uhd::rfnoc::node_t::set_prop_forwarding_map()).
There are several forwarding policies (see uhd::rfnoc::node_t::forwarding_policy_t). The most common forwarding policy is uhd::rfnoc::node_t::forwarding_policy_t::ONE_TO_ONE. Here, properties that are applied to an input edge are forwarded to the corresponding output edge. For more special use cases, uhd::rfnoc::node_t::forwarding_policy_t::USE_MAP can be used to define forwarding rules for non-static cases (e.g., see uhd::rfnoc::switchboard_block_control).
There are some properties that are commonly used, and blocks should use these property names if appropriate:
Edge Properties:
type: It is recommended to always use this property on an edge. It should describe the data type on this edge so the graph can see if types match. The type descriptions are the same as with CPU and OTW types (e.g., sc16, fc32, sc8, u8, ...).samp_rate: Whenever a "sampling rate" is applicable, edges should set this property. DDC, DUC, and Radio blocks all use this for configuring rates and verifying sample rates match.scaling: This property describes how a block has distorted a signal with respect to amplitude. A scaling value of 0.99 means that a signal that used to be fullscale (max. amplitude of 1.0) will only have a maximum amplitude of 0.99. For example, the DDC and DUC blocks use this property to signal how much they have affected the amplitude. The TX and RX streamer objects can then apply the inverse of this scaling factor to correct for the scaling before returning the signal to the user.atomic_item_size: Blocks (e.g. uhd::rfnoc::radio_control) might need a non-dividable amount of data per clock cycle, e.g. the radio block needs sample_per_cycle times the size of the type bytes of data in each cycle. This property allows a block to specify this amount of data per cycle which allows other blocks or the streamer objects to adapt the samples per packet accordingly. Note: Because all blocks in a graph must agree on one value for this property the blocks have to calculate the atomic_item_size as the least common multiple of their own atomic_item_size and the edge property to accommodate for other blocks up- or downstream.User properties:
spp: Blocks that produce data in chunks of samples can use this to describe their "sample per packet" values. Example: uhd::rfnoc::radio_controlinterp, decim: Interpolation and decimation factors when doing rational resampling. Examples: uhd::rfnoc::ddc_block_control, uhd::rfnoc::duc_block_controlSee also the following sections for properties that receive special treatment by the framework.
There is a small number of edge properties that treated differently by the RFNoC framework.
The "tick rate" is the rate that is used for converting floating point timestamps into integer ticks. The tick rate has the additional constraint that within a graph, the tick rate must be the same for all blocks. Therefore, the tick rate is defined by the framework for all blocks that derive from uhd::rfnoc::noc_block_base, and block authors cannot register properties named tick_rate.
The tick rate property is propagated to all edges to enforce all blocks having the same tick rate. Some blocks (like the radio blocks) set the tick rate via their direct access to the hardware, but most blocks use the property propagation mechanism to distribute the tick rate.
To read back the tick rate, use uhd::rfnoc::noc_block_base::get_tick_rate(). Within a block implementation, the uhd::rfnoc::noc_block_base::set_tick_rate() API call can be used to update the tick rate (most blocks do not have to do this).
Because the maximum transmission unit (MTU) is part of the RFNoC framework (it is constrained by the buffer size chosen between blocks, and is thus read back from a register in the FPGA), its property is also created by the RFNoC framework for all blocks, and block authors cannot use the name mtu for their properties.
However, MTUs can differ within a graph, and blocks might have additional constraints on MTU. For this reason, accessing the MTU properties directly is not possible from within a block controller, but rfnoc::noc_block_base provides several APIs to interact with MTUs:
Many blocks (e.g., uhd::rfnoc::ddc_block_control, uhd::rfnoc::duc_block_control) do not resize blocks as they pass through them. That means the input MTU of these blocks equals the output MTU. Thus, they set the MTU forwarding policy to uhd::rfnoc::node_t::forwarding_policy_t::ONE_TO_ONE.
1.8.17