CIRCT

Circuit IR Compilers and Tools

ESI data types and communication types

ESI has two different classes of MLIR types: ones which represent data on the wires (data types) and ones which specify the type of communication. From a user perspective, communication types aren’t really types – this is just how the communication style is modeled in MLIR and thus an implementation detail.

Data types 

In addition to the types in the hw dialect, ESI will add few:

Void 

void translates to “no data”, meaning just a control signal.

Status: planned

Lists 

Lists are used to reason about variably-sized data. There are two types of lists: those for which the size is known before transmission begins (a fixed size list, or fixed_list), and those for which it isn’t (a variably sized list, or just list). fixed_lists are very roughly similar to a C pointer/size pair (in that they contain a size upfront) whereas lists are roughly similar to a simply singlely linked list wherein the size is not known until one iterates through it.

Since it isn’t yet clear what (if any) benefit fixed_list will provide, it is not plan-of-record to implement it. list will be implemented in the future.

Status: planned

Communication types 

As stated above, communication “types” do not represent data on the wire; rather, they speak to the signaling.

Channels 

ESI “channels” are streaming connections upon which messages can be sent and received. They are expressed by wrapping the type (e.g. !esi.channel<i5>) and using it like any other value type.

hw.module @add11(%clk: i1, %ints: !esi.channel<i32>) -> (%outs: !esi.channel<i32> ) {
  hw.output %ints
}

The value which the channel is carrying must be (un)wrapped to access it:

hw.module @add11(%clk: i1, %ints: !esi.channel<i32>) -> (mutatedInts: !esi.channel<i32>) {
  %i, %i_valid = esi.unwrap.vr %ints, %rdy
  %c11 = hw.constant 11
  %m = comb.add %c11, %i
  %mutInts, %rdy = esi.wrap.vr %m, %i_valid
  hw.output %mutInts
}

[insert diagram of wrap/unwrap]

Status: complete and stable (for supported data types)

Memory-mapped IO 

This section is not fully thought out or written about. It is certainly not implemented. The text in this section should be considered initial thoughts.

Status: Planning

MMIO Regions 

The basic idea is to present ESI memory mapped regions exposed by modules. There could be any number of these regions exposed and they would work somewhat like input (request) / output (response) port pairs, but with implicit request-response signaling and structure. The MMIO space itself would be defined by a statically sized struct (so lists would be disallowed), with address offsets implicitly or explicitly defined. The method for base address assignment is yet to be decided. These regions can support atomic reads or writes of arbitrary size or limited size.

MMIO Requests (read/write) 

Other modules connected to the same MMIO bus could specify read/write requests in several ways:

  • As a data window with blinds (or similar construct) specifying the struct fields to read/write
  • As a list of address offsets and sizes or data to read/write

Along with the request content, atomicity of the request would have to be specified. The response on a read would correspond to the read request – either a list of bytes read or a data window with the requested data filled in. For a write, a simple acknowledgement or error would suffice for the response.

Automatic self-description 

An MMIO region’s data type can be used to automatically generate a self-describing software data type and an access API. Additionally, a per MMIO bus table could be generated with the base addresses for each connected region and a descriptor for each.

Software access 

The MMIO struct would become a software struct which software could map to the base address as a typed pointer. This way, software access would be simpler (just a normal struct pointer dereference – e.g. p->field1) and safer as it knows the correct address implicitly and the type of that field. Additionally, if the MMIO region knows the processor’s endianness, it could respond in the correct endianness. How to initiate an atomic read/write of multiple fields from software is undecided yet, though there may be some merit to exposing data windows to software.

Additional Type: any 

An MMIO space can be have type any to allow access to memory spaces which are fundamentally untyped (e.g. memory-mapped host RAM). In this case, the requestor can specify the type. For instance, to access 64-bit values the requestor would send a request with a given address and specify a uint<64> type response or write. Alternatively, it could request a response or write type of type struct ConfigSpace to specify it wants to read or write a configuration.

Implementation 

The MMIO system could be implemented on top of the streaming portion.

Data windows

By default, an entire message is guaranteed to be presented to the receiving module in one clock cycle (excepting lists). For particularly large messages, this is not ideal as it requires a data path equal to the size of the message. Data windows specify which parts of a message a module accepts on each clock cycle. For structs, they specify which members are accepted on which cycles. For arrays and lists, they specify how many items can be accepted each clock cycle.

Data windows do not affect port compatibility. In other words, a designer can connect ports using two different windows into the same data type. This can be used to connect different modules with different bandwidths. A data window merely specifies the logical “gasket” used to connect two differently sized ports.

The MLIR representation of data windows has yet to be determined.

Status: planning