RISu64 (Reduced Instruction Set μProcessor 64 / Squirrel 64) is a series of my toy 64-bit RISC-V compatible processors. RISu064 (this repo) is the first in the series. Illustration by Andy Lithia.
- RV64IMZicsr_Zifencei instruction set
- 7-stage pipeline: PCGen(F1), IMem(F2), Decode(ID), Issue(IX), Execute(EX), DMem(MEM), Writeback(WB).
- In-order issue and out-of-order writeback
- Dual-issue
- BTB + Bimodal/Gselect/Gshare/Tournament + RAS branch predictors
- 2x Integer (arithmetic, barrel shifter, branch)
- 1x Load store unit (aligned access only, unaligned access generate precise exception)
- 1x Multiply/ divide unit (non-pipelined, 3/6-cycle 32/64bit multiply, 34/66-cycle 64bit divide)
- Multiply/ divide is optional
- Optional L1 instruction and data cache (2-way set associative blocking cache)
- Machine mode with exception and interrupt support
- Optional experimental hardware refilled MMU + supervisor and user mode support
- Written in portable synthesizable Verilog
The performance varies based on configurations:
- Single-issue + 512-entry Bimodal + 32-entry BTB + TCM: 3.01 Coremark/MHz
- Single-issue + 4K-entry Tournament + 32-entry BTB + TCM: 3.06 Coremark/MHz
- Single-issue + 4K-entry Tournament + 32-entry BTB + 16KB L1$: 3.01 Coremark/MHz
- Dual-issue + 4K-entry Tournament + 32-entry BTB + TCM: 4.31 Coremark/MHz
Note:
- Compiled with GCC 9.2.0, with the following options:
-MD -O3 -mabi=lp64 -march=rv64im -mcmodel=medany -ffreestanding -nostdlib -fomit-frame-pointer -funroll-all-loops -finline-limit=1000 -ftree-dominator-opts -fno-if-conversion2 -fselective-scheduling -fno-code-hoisting -freorder-blocks-and-partition - Single-issue is no longer supported in the latest branch, testing was carried out using commit
efd0d3 - L1-cache is organized as 2-way set associative, 16KB each, with simulated unlimited L2 memory and 15-cycle latency
- Each BPU entry is 2-bit, internally it expects 8-bit wide memory interface. 4K-entry = 1K x 8bit SRAM
The area is quite big right now (rather poor PPA).
FPGA:
Currently the multiplier is not optimized for FPGA yet. With Aritx-7 XC7A100T-3CSG324C:
- Multiplier disabled, no cache: ~120 MHz fmax, 19.6K LUT, 6.9K FF
The critical path is at write-back stage.
ASIC:
The project has been submitted to Google + efabless MPW-7 shuttle for tapeout, with a 5GHz narrow-band RF transceiver.
The total area allocated to this project is about 8.5mm^2. The core is configured to be:
- 4K depth Gshare predictor
- 8KB 2-way I-cache + 8KB 2-way D-cache
- Hardware multiplier and divider enabled
- MMU disabled, machine mode only
Total area allocated to core minus SRAM cell is about 3.4mm^2, with around 39% utilization. Assuming 85% target placement density, this translate to a 1.56mm^2 die area at SKY130 process with SKY130HD cell library.
Regarding maximum frequency, without SRAM/ cache, Fmax is around 100MHz with CLA+KSA hybrid adder, or 80MHz with inferred adder. With cache, tag comparsion logic becomes the critical path and Fmax drops to about 50MHz.
This project is mostly a proof-of-concept and is regarded as done. There might be bug fixes in the future, but don't expect major changes.
In sim folder, run make. It should build the simulator.
To run coremark, build the coremark by running make in tests/coremark, then in the sim folder do ./simulator --ram ../tests/coremark/coremark.bin.
Note: Verilator required for building the simulator. RV64 gcc (riscv64-unknown-elf-gcc) required for building the coremark.
The core implementation probably contains bugs. Due to its OoO WB without reordering design, the core's architectural state would often diverge from ISA model, making lock-step co-simulation or trace comparsion with ISA simulation hard. A trace comparison tool is provided to allow comparing between RTL simulator generated trace and Spike generated trace. Example usage:
spike -m0x20000000:4096,0x80000000:1048576 -l --log-commits tests/coremark/coremark.elf 2> spike.log
sim/simulator --ram tests/coremark/coremark.bin --cycles 10000 > sim.log
tests/trace_comparater.py --risu sim.log --spike spike.log
Differences (if any) will be reported.
During the design of this processor, I have used the following projects as reference:
- lowRISC's muntjac, Apache 2.0 license
- UltraEmbedded's biriscv, Apache 2.0 license
The following third-party code have been used:
- Gary Guo's round robin arbiter, BSD license
MIT