Minerva

A 32-bit RISC-V soft processor

Minerva is a CPU core that currently implements the RISC-V RV32IM instruction set. Its microarchitecture is described in plain Python code using the nMigen toolbox.

Quick start

Minerva currently requires Python 3.6+ and nMigen on its master branch.

python setup.py install
python build.py > minerva.v

To use Minerva, you need to wire the following ports to minerva_cpu:

• clk
• rst
• ibus__*
• dbus__*
• external_interrupt

Features

The microarchitecture of Minerva is largely inspired by the LatticeMico32 processor.

Minerva is pipelined on 6 stages:

1. Address The address of the next instruction is calculated and sent to the instruction cache.
2. Fetch The instruction is read from memory.
3. Decode The instruction is decoded, and operands are either fetched from the register file or bypassed from the pipeline. Branches are predicted by the static branch predictor.
4. Execute Simple instructions such as arithmetic and logical operations are completed at this stage.
5. Memory More complicated instructions such as loads, stores and shifts require a second execution stage.
6. Writeback Results produced by the instructions are written back to the register file.

The two L1 caches are write-through, meaning that writes are done to both the cache and the underlying memory hierarchy.

Miss penalties are reduced by restarting execution as soon as the missing word is available, without waiting for the complete line refill. Furthermore, cache refills always request the missing word first. By combining these two policies, Minerva is able to resume its execution after a cache miss in only two clock cycles in the best case, regardless of line size.

The L1 data cache is coupled to a write buffer. Store transactions are in this case done to the write buffer instead of the data bus. This enables stores to proceed in one clock cycle if the buffer isn't full, without having to wait for the bus transaction to complete. Store transactions are then completed in the background as the write buffer gets emptied to the data bus.

Minerva is able to operate at 100MHz on a Xilinx XC7A35-2 FPGA, with both caches enabled.

Configuration

The following parameters can be used to configure the Minerva core.

Parameter	Default value	Description
reset_address	0x00000000	Reset vector address
with_icache	True	Enable the instruction cache
icache_nb_ways	1	Number of ways in the instruction cache
icache_nb_lines	256	Number of lines in the instruction cache
icache_nb_words	8	Number of words in a line of the instruction cache
icache_base	0x00000000	Base of the instruction cache address space
icache_limit	0x80000000	Limit of the instruction cache address space
with_dcache	True	Enable the data cache
dcache_nb_ways	1	Number of ways in the data cache
dcache_nb_lines	256	Number of lines in the data cache
dcache_nb_words	8	Number of words in a line of the data cache
dcache_base	0x00000000	Base of the data cache address space
dcache_limit	0x80000000	Limit of the data cache address space
as_instance	False	Add a default clock domain
with_muldiv	False	Enable RV32M support
with_debug	False	Enable the Debug Module
with_trigger	False	Enable the Trigger Module

Possible improvements

In no particular order:

• RV64I
• Floating Point Unit
• Stateful branch prediction
• ...

If you are interested in sponsoring new features or improvements, get in touch at contact [at] lambdaconcept.com .

License

Minerva is released under the permissive two-clause BSD license. See LICENSE file for full copyright and license information.