Skip to content

Latest commit

 

History

History
298 lines (206 loc) · 14.6 KB

README.md

File metadata and controls

298 lines (206 loc) · 14.6 KB
Raw

Spacing

FazyRV -- A Scalable RISC-V Core

A minimal-area RISC-V core with a scalable data path to 1, 2, 4, or 8 bits and manifold variants.

Table of Content

Introduction

FazyRV is a minimal-area RISC-V RV32 core with inherent scalability. The data path can be set to a width of either 1, 2, 4, or 8 bits to process smaller chunks of the operands each clock cycle. Scaling the chunk size allows a trade-off between area and performance at synthesis time. Moreover, each chunk size can be combined with manifold variants to find the best-fitting configuration and trade-off for given system requirements and technology. In contrast to other approaches, FazyRV tries to avoid manual optimization at the gate level (see also Decoder).

Area / Resource Demand

The plot below tracks the resource demand of the FazyRV core (left) and a minimal reference SoC (right) for an iCE40 architecture. Note that only a few variants are plotted for brevity.

FazyRV and fsoc areas

Organization

FazyRV is contained in rtl and licensed under a permissive MIT license. rtl/fazyrv_top.sv is the top module and instantiates the FazyRV core (rtl/fazyrv_core.sv) alongside the register file. The reference SoC (fsoc) is located in soc/rtl along with a constraint file for common architectures.

The data flow through the core can be outlined as follows:

block diagram

Acknowledgement

TU Graz Logo Spacing EAS Logo

This repository includes results of research in the Embedded Architectures & Systems (EAS) Group at Graz University of Technology.

 

Design Variants

CHUNKSIZE sets the data path width of FazyRV and, thus, primarily determines the required Cycles per Instruction (CPI). When the chunk size is set to 1, a single bit is processed per clock cycle, and an addition, e.g., requires 32 cycles (without fetch and decode) to process the operands. It reduces to 4 clock cycles when 8 bits are processed simultaneously, i.e., CHUNKSIZE = 8. The latter variant, however, comes with an increased area demand.

CHUNKSIZE   := 1 | 2 | 4 | 8

The configuration (CONF) determines the functionality of the core. In the MIN variant, no interrupts are supported and the core implements to the minimal area achievable. The work-in-progress INT variant adds basic interrupt support but sacrifices many features in favor of reduced area. Thus, e.g., mtvec can only be set at synthesis time. Finally, the work-in-progress CSR variant adds more CSR registers and allows reading and writing them. However, there are deviations from the specification as area remains a main optimization objective.

CONF        := MIN | INT | CSR

The register file type (RFTYPE) influences both the area and the performance. It selects how registers are implemented in FazyRV. The most area-intensive option is LOGIC. It implements the registers in logic elements such as LUT RAM or FF RAM. Due to the large area, this option is primarily used for testing purposes and cannot be combined with all variants. BRAM implements the register file in block RAM with a single read port. Thus, operands must be fetched sequentially, leading to a performance decrease compared to a dual-port RAM implementation (BRAM_DP). The BP variants BRAM_BP and BRAM_DP_BP additionally instantiate a bypass multiplexer to make the operands accessible to the core one cycle in advance. While this improves the performance due to the saved clock cycle per instruction, more area may be consumed.

RFTYPE      := LOGIC | BRAM | BRAM_BP | BRAM_DP | BRAM_DP_BP

In the default configuration, FazyRV interfaces the memory with a Wishbone interface. MEMDLY1 allows deviation from the handshake and assumes that each access is directly granted with a one-clock cycle delay until the access is completed and the results are available. This can lower the average CPI when the design ensures the above conditions are met, e.g., when the memory is implemented in BRAM.

MEMDLY1     := 0 | 1

Important

Please note that the core still needs to become production-ready. As there is a strong focus on area, the implemented features in the INT and CSR variants and allowed deviations from the specification must be carefully considered. Also, note that the LOGIC variant is explicitly not considered with INT and CSR variants.

Quick Start

Prepare the Environment

You may want to use a Python virtual environment.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Reference SoC

fsoc (soc/fsoc) is a minimal SoC used to track the core's area demand and run simulation-based tests and benchmarks. The Make targets for the reference implementations are based fusesoc (for most FPGA architectures). fazyrv.core and fsoc.core core files for FazyRV and fsoc, respectively.

Start by adding core files to the fusesoc library.

fusesoc library add fazyrv .
fusesoc library add fsoc .

Then, you can use the targets in the core file or run the flow with the Make targets below. A reference implementation for the following architectures is supported. Replace <ARCH> with the desired target architecture (see below). Also, replace <CHUNKSIZE>, <CONF>, and <RFTYPE> with the desired values.

ARCH        := ice40 | ecp5 | gowin | xilinx | gatemate

# run the flow only
make _impl.soc.<ARCH>-<CHUNKSIZE>-<CONF>-<RFTYPE>

# run the flow and report a summary of the results
make _report.soc.<ARCH>-<CHUNKSIZE>-<CONF>-<RFTYPE>

# e.g.,
make _report.soc.ice40-8-MIN-BRAM

# or 
make report.soc.all

Litex

LiteX supports FazyRV with the following options: --cpu-chunksize to set the chunk size (1, 2, 4, or 8), --cpu-conf to set the configuration (MIN, INT, or CSR) and --cpu-rftype to set the register file type (LOGIC, BRAM, BRAM_BP, BRAM_DP, or BRAM_DP_BP).

litex_sim --cpu-type=fazyrv --cpu-chunksize=4 --cpu-rftype=LOGIC

Tests and Verification

Run riscv-tests

riscv-tests are used as fast checks to get feedback if a variant is broken. However, the tests are not as sensitive as the RISCOF tests and may report false positives.

Execute the tests with the make target below by replacing the <CHUNKSIZE>, <CONF>, and <RFTYPE> with the desired variant or run it on all variants.

# riscv-tests
make sim.riscvtests.<CHUNKSIZE>-<CONF>-<RFTYPE>
# e.g.,
make sim.riscvtests.8-MIN-BRAM
# or
make report.riscvtests.all

Run RISCOF

The RISCOF framework provides more extensive simulation-based design tests. You can run the test either for one variant or use riscof.all to run the tests for a selected subset of all variants.

# RISCOF
make riscof.prepare
make riscof.run.<CHUNKSIZE>-<CONF>-<RFTYPE>
# e.g.,
make riscof.prepare
make riscof.run.8-MIN-BRAM
# or
make riscof.prepare
make riscof.all

Module-Level Formal Checks

The ALU (rtl/fazyrv_alu.sv) and the spm_d module (rtl/fazyrv_spm_d.sv) are checked by BMC in a formal test bench. The test benches and .sby files are located in fv/alu and fv/spm_d, respectively. These were primarily used to support a formal verification test-driven development when the core was not ready to be checked by riscv-formal. However, they remain important to verify changes and optimizations. The chunk size is set to 8 by default. If required, please update the local parameter parameter CHUNKSIZE in the formal test benches accordingly.

# alu
cd fv/alu
sby -f fazyrv_alu_bmc.sby
sby -f fazyrv_alu_cov.sby

# spm_d
cd fv/spm_d
sby -f fazyrv_spm_d_bmc.sby
sby -f fazyrv_spm_d_cov.sby

riscv-formal

In addition to simulation-based tests, formal checks are applied using riscv-formal. Due to the exponential run time, formal checks are primarily considered for larger chunk sizes. Also, the depth is limited.

# insn checks
make fv.rvformal.bmc.insn.<CHUNKSIZE>
make fv.rvformal.cov.insn.<CHUNKSIZE>

# reg checks
make fv.rvformal.bmc.reg.<CHUNKSIZE>
make fv.rvformal.cov.reg.<CHUNKSIZE>

# e.g.
make fv.rvformal.bmc.insn.8 && make fv.rvformal.cov.insn.8

# or run them sequentiall on all chunk sizes
make fv.rvformal.bmc.insn.al
make fv.rvformal.bmc.reg.all
make fv.rvformal.cov.insn.all
make fv.rvformal.cov.reg.all

Benchmarks

Embench

Embench is used to benchmark FazyRV and compare it to similar minimal-area cores. It can be run via a Make target.

make embench.run

The target calls the shell script script/benchmark_run_embench_all.sh. Please adapt it to run the benchmark suite on the desired variants. --insn_timing is used to store information about all executed instructions on the disk. It can be used to analyze and compare the cycles per instructions (CPI). Note that this significantly increases the required disk space.

Decoder

FazyRV is written at a Verilog RTL abstraction level that remains human-readable. Thus, hand-optimization at the gate level is avoided whenever possible. However, some parts of the design benefit from optimization at a lower level, such as the combinational logic in the instruction decoder. Thus, the decoder is given in a table-like format that can be fed into the ESPRESSO logic optimizer to generate an optimized gate-level description. The input file (optimizer/decoder/decoder) can be edited efficiently. The instruction bits can either be checked explicitly or set to don't care when they are not required to interpret legal instructions. The former has more freedom to optimize to a smaller implementation but is more insecure when illegal instructions appear.

espresso -o eqntott -Dso_both ~/Documents/public_repos/FazyRV/optimizer/decoder/decoder > tmp

The output gate-level Verilog code can then be updated in the decoder implementation (rtl/fazyrv_decode.sv).

A simple fuzzing script is used to estimate the optimization potential of the decoder. It iteratively sets bits to don't care and tests the modification. First, riscv-tests are run to find broken decoders quickly and shorten the overall run time. When the riscv-tests pass, a riscv-formal formal verification run is started on the insn checks for the 8-bit variant. Then, the modification is either reverted or the don't care is kept. The fuzzer is solely for testing purposes. It is not recommended to take over the optimized result. The final decoder must be carefully checked and thoroughly verified on all variants before using it in the core. Please note that the current version only tests one bit at a time and may be improved.

cd optimizer/decoder/decode_opt
python3 fuzz.py --espresso_file ../decoder --riscvtests_dir ../../../sim --riscvformal_dir ../../../ --template_verilog fazyrv_decode.template --template_marker "//<PUT_IT_HERE>" --destination_verilog ../../../rtl/fazyrv_decode.sv --espresso_optimized espresso.optimized

Related Resources and Further Readings

  • We presented this work at the 21st ACM International Conference on Computing Frontiers (CF '24). You can find the paper here (open access). It summarizes our design objectives, gives insight into the design and trade-offs, compares similar cores, and provides an in-depth evaluation.

Please cite the work as follows:

@inproceedings{fazyrv2024kissich,
  title = {{FazyRV: Closing the Gap between 32-Bit and Bit-Serial RISC-V Cores with a Scalable Implementation}},
  booktitle = {Proc. of the 21st ACM International Conference on Computing Frontiers (CF ’24)},
  author = {Kissich, Meinhard and Baunach, Marcel},
  year = {2024},
  month = {May},
  publisher = {Association for Computing Machinery},
  url = {https://doi.org/10.1145/3649153.3649195},
  doi = {10.1145/3649153.3649195},
  booktitle = {Proceedings of the 21st ACM International Conference on Computing Frontiers},
  pages = {240–248}
}

TODOs

  • Workflow: caching, tool versions, artifacts, dependence on some local tools
  • RVC extension (compressed instructions)
  • INT variant
  • CSR variant
  • CSR instructions in addition to csrrw(?)
  • Use edalize reporting instead of custom scripts
  • Optimization
  • More documentation

Please feel free to discuss and open an issue and/or pull request.

Licensing

The FazyRV core (rtl/*) is licensed under the MIT license. This license may not apply to the remainder of the repository.