riscv-rust-baremetal

Rust example that compiles to baremetal riscv. Supports HTIF to communicate with host and perform proxy syscalls.

Quick start

Edit the configuration at .cargo/config.toml.

rustup target add riscv64gc-unknown-none-elf
cargo build

To run a specific benchmark (Requires Spike):

cargo run --bin bin-name

You can also run cargo with specific configuration with the --config flag. Example:

rustup target add riscv32imac-unknown-none-elf
cargo --config 'build.target="riscv32imac-unknown-none-elf"' \
    --config 'target.riscv32imac-unknown-none-elf.linker="riscv32-unknown-elf-ld"' \
    r --bin median --release

TODO

• sync syscall (https://stackoverflow.com/questions/72369202/are-mutable-static-primitives-actually-unsafe-if-single-threaded)
  • spin crate provides spinlock, or just write our own
  • https://matklad.github.io/2020/01/02/spinlocks-considered-harmful.html
• make work in release mode
• exit properly
• core::write
• trap handling
• panic handling
• double check qsort

Ideas

• Unsafe rust (get_unchecked, unwrap_unchecked)
• Idiomatic rust using iterators
• Share rust code/interface with fesvr for syscalls
• Would be cool to debug with gdb using openocd: https://docs.rust-embedded.org/book/start/hardware.html

Benchmarks

• Port benchmarks from riscv tests

• Cycles should be similar?

• Not sure if cycles are accurate, might need more testing / investigating.

• Figured out why mcycle is worst: no vector instructions

• Tower: The the original riscv-tests uses linked list. Linked list normally sucks, but maybe in embedded programming it is good. Still need to decide how to port

• memcpy: clone (memcpy) uses 5x more mcycle, maybe look at rust documentation. C version unrolls the loop

• radix sort: https://crates.io/crates/rdx/0.0.7

Conclusion

Hard to recreate some of the tests without using unsafe and pointers, but then we would just be writing C. Maybe we should still do that.

Borrow checker is pretty good. Safety gaurantees are nice too. For example, global static is unsafe by default, so the programmer knows that they need sync or redesign (for porting riscv-tests, I make the caller track data for now instead of global statics).

Results

Benchmark	Rust mcycle	C mcycle	Godbolt
median	6449	4654	https://godbolt.org/z/8rdThb57a
sort	215018	x
qsort	231015	123208
qsort (with no bounds checking)	184037
rsort	TODO	171156
towers	TODO
vvadd	4835	2411	https://godbolt.org/z/e1jcPajGW
memcpy	9540	5515	https://godbolt.org/z/Wz3T6oP96
multiply	7590	26811	https://godbolt.org/z/6fjjqP6rq
multiply (with --unroll-loops)		20986	https://godbolt.org/z/6fjjqP6rq

Resources

HTIF implementations:

• https://github.com/ucb-bar/libgloss-htif
• https://github.com/riscv-software-src/riscv-tests/tree/master/benchmarks/common

Rust:

• Baremetal rust: https://docs.rust-embedded.org/book/intro/index.html
• Riscv rust: https://github.com/rust-embedded/riscv