Circuit IR Compilers and Tools

Miscellaneous Notes


The ABI specifies how an ESI “API” is translated to hardware, specifically RTL. This involves both the wire-level signaling between modules and how data is arranged on those wires.

This section is purposely underspecified in this proposal as it should be an implementation detail which only advanced users need know. The main issues discussed here are how lists and data windows are lowered and presented to RTL modules. Lowering of fixed-size, default-presentation semantics ports is mostly straight forward, so is not discussed here.

Wire-level signaling 

Since ESI connections are elastic (latency-insensitive), the signaling scheme must include some notion of validity and some notion of backpressure. This rules out interfaces which inherently cannot be back-pressured, though future ESI specs may support non-backpressure (feed-forward) interfaces. For now, a buffering gasket which converts between ESI and the non-backpressure-able interface is necessary, perhaps implementing backpressure in a higher-level protocol (e.g. not initiating DMA transfers until space exists to buffer them). These typically exist somewhere in many if not most designs.

It is up to the compiler to determine the appropriate wire signaling scheme to use for message transfer between ESI interfaces. (The designer can optionally specify it manually.) If buffering is required, the module should state that it is required and then the compiler should build it automatically. The interface to an RTL module, however, is up to that particular module.

There are several existing standards which we should consider implementing:

  • AXI Stream/AMBA
  • Avalon-MM/Avalon-ST
  • Simple valid/ready-ack semantics (this should be the default) for streaming
  • Simple RW RAM-style interface (this should be the default) for MMIO

Lists and Data Windows 

The basic idea behind presenting lists and data windows to modules (at least with simple valid/ready-ack semantics) is that ESI can automatically construct the proper state machines automatically and to essentially present a discriminated union which informs the RTL which part of the message is currently being presented – the particular data window part and/or position in a list.

A note on language implementation 

It is not always clear how a particular hardware design language (HDL) would be made to support all of the ESI constructs. That’s fine – not all HDLs are suited to all the ESI constructs. For instance, some languages are only designed for data stream processing so only the data channel (streaming) parts of ESI make sense. Maybe a compiler for said language would also generate some MMIO regions/clients (to access DRAM, PCIe, network, et cetera).

An HDL compiler which supports ESI is not required to implement the entire ESI spec. Rather, it is encouraged to only implement the parts of ESI that make sense; however, it is also encouraged that the compiler authors think long and hard about which constructs make sense. For example, for an RTL compiler it may not be immediately obvious how to support lists, the variable length data type. After further consideration, however, one may realize that RTL can accept variable length data over multiple cycles. In fact, this is the intention as variable length data in hardware is generally reasoned about and implemented in the temporal dimension rather than the spatial one. “General purpose” HDLs (i.e. SystemVerilog) compilers are encouraged to support all of the ESI specification.