This repository provides modules to build on-chip communication networks adhering to the AXI4 or AXI4-Lite standards. For high-performance communication, we implement AXI4+ATOPs from AXI5. For lightweight communication, we implement AXI4-Lite. We aim to provide a complete end-to-end communication platform, including endpoints such as DMA engines and on-chip memory controllers.
Our design goals are:
- Topology Independence: We provide elementary building blocks such as protocol multiplexers and demultiplexers that allow users to implement any network topology. We also provide commonly used interconnecting components such as a crossbar.
- Modularity: We favor design by composition over design by configuration where possible. We strive to apply the Unix philosophy to hardware: make each module do one thing well. This means you will more often instantiate our modules back-to-back than change a parameter value to build more specialized networks.
- Fit for Heterogeneous Networks: Our modules are parametrizable in terms of data width and transaction concurrency. This allows to create optimized networks for a wide range of performance (e.g., bandwidth, concurrency, timing), power, and area requirements. We provide modules such as data width converters and ID width converters that allow to join subnetworks with different properties, creating heterogeneous on-chip networks.
- Full AXI Standard Compliance.
- Compatibility with a wide range of (recent versions of) EDA tools and implementation in standardized synthesizable SystemVerilog.
The design and microarchitecture of the modules in this repository is described in this paper (preprint). If you use our work in your research, please cite it.
In addition to the documents linked in the following table, we are setting up documentation auto-generated from inline docstrings. (Replace master in that URL with a tag to get the documentation for a specific version.)
| Name | Description | Doc |
|---|---|---|
axi_atop_filter |
Filters atomic operations (ATOPs), i.e., write transactions that have a non-zero aw_atop value. |
|
axi_burst_splitter |
Split AXI4 burst transfers into single-beat transactions. | |
axi_cdc |
AXI clock domain crossing based on a Gray FIFO implementation. | |
axi_cut |
Breaks all combinatorial paths between its input and output. | |
axi_delayer |
Synthesizable module which can (randomly) delays AXI channels. | |
axi_demux |
Demultiplexes an AXI bus from one slave port to multiple master ports. | Doc |
axi_dw_converter |
A data width converter between AXI interfaces of any data width. | |
axi_dw_downsizer |
A data width converter between a wide AXI master and a narrower AXI slave. | |
axi_dw_upsizer |
A data width converter between a narrow AXI master and a wider AXI slave. | |
axi_err_slv |
Always responds with an AXI decode/slave error for transactions which are sent to it. | |
axi_fifo |
A Fifo for each AXI4 channel to buffer requests. | |
axi_from_mem |
This module acts like an SRAM and makes AXI4 requests downstream. | |
axi_id_prepend |
This module prepends/strips the MSB from the AXI IDs. | |
axi_id_remap |
Remap AXI IDs from wide IDs at the slave port to narrower IDs at the master port. | Doc |
axi_id_serialize |
Reduce AXI IDs by serializing transactions when necessary. | Doc |
axi_intf |
This file defines the interfaces we support. | |
axi_isolate |
A module that can isolate downstream slaves from receiving new AXI4 transactions. | |
axi_iw_converter |
Convert between any two AXI ID widths. | Doc |
axi_join |
A connector that joins two AXI interfaces. | |
axi_lfsr |
AXI4-attached LFSR; read returns pseudo-random data, writes are compressed into a checksum. | |
axi_lite_demux |
Demultiplexes an AXI4-Lite bus from one slave port to multiple master ports. | Doc |
axi_lite_dw_converter |
A data width converter between two AXI-Lite busses | [Doc][doc.axi_lite_dw_converter] |
axi_lite_from_mem |
This module acts like an SRAM and makes AXI4-Lite requests downstream. | |
axi_lite_join |
A connector that joins two AXI-Lite interfaces. | |
axi_lite_lfsr |
AXI4-Lite-attached LFSR; read returns pseudo-random data, writes are compressed into a checksum. | |
axi_lite_mailbox |
A AXI4-Lite Mailbox with two slave ports and usage triggered irq. | Doc |
axi_lite_mux |
Multiplexes AXI4-Lite slave ports down to one master port. | Doc |
axi_lite_regs |
AXI4-Lite registers with optional read-only and protection features. | Doc |
axi_lite_to_apb |
AXI4-Lite to APB4 protocol converter. | |
axi_lite_to_axi |
AXI4-Lite to AXI4 protocol converter. | |
axi_lite_xbar |
Fully-connected AXI4-Lite crossbar with an arbitrary number of slave and master ports. | Doc |
axi_modify_address |
A connector that allows addresses of AXI requests to be changed. | |
axi_multicut |
AXI register which can be used to relax timing pressure on long AXI buses. | |
axi_mux |
Multiplexes the AXI4 slave ports down to one master port. | Doc |
axi_pkg |
Contains AXI definitions, common structs, and useful helper functions. | |
axi_rw_join |
Joins a read and a write slave into one single read / write master. | |
axi_rw_split |
Splits a single read / write slave into one read and one write master. | |
axi_serializer |
Serializes transactions with different IDs to the same ID. | |
axi_throttle |
Limits the maximum number of outstanding transfers sent to the downstream logic. | |
axi_test |
A set of testbench utilities for AXI interfaces. | |
axi_to_axi_lite |
AXI4 to AXI4-Lite protocol converter. | |
axi_to_mem |
AXI4 to memory protocol (req, gnt, rvalid) converter. Additional banked, interleaved, split variant. | |
axi_xbar |
Fully-connected AXI4+ATOP crossbar with an arbitrary number of slave and master ports. | Doc |
axi_xp |
AXI Crosspoint (XP) with homomorphous slave and master ports. |
The following modules are meant to be used for verification purposes only but are synthesizable to be used in FPGA environments.
| Name | Description |
|---|---|
axi_bus_compare |
Compares two buses of the same type (and in the same clock domain), returns events on mismatch. |
axi_slave_compare |
Compares two slave devices of the same type (and in the same clock domain), returns events on mismatch. |
In addition to the modules above, which are available in synthesis and simulation, the following modules are available only in simulation. Those modules are widely used in our testbenches, but they are also suitable to build testbenches for AXI modules and systems outside this repository.
| Name | Description |
|---|---|
axi_chan_compare |
Non-synthesizable module comparing two AXI channels of the same type |
axi_chan_logger |
Logs the transactions of an AXI4(+ATOPs) port to files. |
axi_driver |
Low-level driver for AXI4(+ATOPs) that can send and receive individual beats on any channel. |
axi_dumper |
Dumps log to file to be interpreted by axi_dumper_interpret script for debugging purposes. |
axi_lite_driver |
Low-level driver for AXI4-Lite that can send and receive individual beats on any channel. |
axi_lite_rand_master |
AXI4-Lite master component that issues random transactions within user-defined constraints. |
axi_lite_rand_slave |
AXI4-Lite slave component that responds to transactions with constrainable random delays and data. |
axi_rand_master |
AXI4(+ATOPs) master component that issues random transactions within user-defined constraints. |
axi_rand_slave |
AXI4(+ATOPs) slave component that responds to transactions with constrainable random delays and data. |
axi_scoreboard |
Scoreboard that models a memory that only gets changed by the monitored AXI4(+ATOPs) port. |
axi_sim_mem |
Infinite memory with AXI4 slave port. |
AXI4+ATOPs means the full AXI4 specification plus atomic operations (ATOPs) as defined in Section E1.1 of the AMBA 5 specification. This has the following implications for modules that do not implement ATOPs and systems that include such modules:
- Masters that do not issue ATOPs must set
aw_atopto'0. - Slaves that do not support ATOPs must specify this in their interface documentation and can ignore the
aw_atopsignal. - System designers are responsible for ensuring that
- slaves that do not support ATOPs are behind an
axi_atop_filterif any master could issue an ATOP to such slaves and - the
aw_atopsignal is well-defined at the input of any (non-AXI4-Lite) module in this repository.
- slaves that do not support ATOPs are behind an
Masters and slaves that do support ATOPs must adhere to Section E1.1 of the AMBA 5 specification. In particular:
- ATOPs that have the
aw_atop[axi_pkg::ATOP_R_RESP]bit set generate a write response (B channel) beat and at least one read response (R channel) beat. All modules for which theaw_atop[axi_pkg::ATOP_R_RESP]bit could be set at their master port must be able to handle both B and R beats (in any order and without requiring a simultaneous handshake) for each such ATOP request. All modules for which theaw_atop[axi_pkg::ATOP_R_RESP]bit could be set at their slave port must respond with the appropriate number of B and R beats for each such ATOP request. - ATOPs must not use the same AXI ID as any other transaction that is outstanding at the same time.
Our code is written in standard SystemVerilog (IEEE 1800-2012, to be precise), so the more important question is: Which subset of SystemVerilog does your EDA tool support?
We aim to be compatible with a wide range of EDA tools. For this reason, we strive to use as simple language constructs as possible, especially for our synthesizable modules. We encourage contributions that further simplify our code to make it compatible with even more EDA tools. We also welcome contributions that work around problems that specific EDA tools may have with our code, as long as:
- the EDA tool is reasonably widely used,
- recent versions of the EDA tool are affected,
- the workaround does not break functionality in other tools, and
- the workaround does not significantly complicate code or add maintenance overhead.
In addition, we suggest to report issues with the SystemVerilog language support directly to the EDA vendor. Our code is fully open and can / should be shared with the EDA vendor as a testcase for any language problem encountered.
All code in each release and on the default branch is tested on a recent version of at least one industry-standard RTL simulator and synthesizer. You can examine the CI settings to find out which version of which tool we are running.