Skip to content

Raven v1.29.0 — many cores, one truth

Choose a tag to compare

@github-actions github-actions released this 03 Nov 18:39
· 8 commits to main since this release

🐦‍⬛ Raven v1.29.0: many cores, one truth

🧭 TL;DR. Each core can now get its own L1 cache, and you pick how those caches stay in step: 🌫️ none, 🟦 MSI, 🟩 MESI or 🟧 MOESI. Every line state, bus request and invalidation shows up live. Underneath that, Raven became a multi-architecture engine 🧬 with four pipeline models 🚦, an RV32F that follows IEEE 754 🧮, and the official riscv-tests suite passing 71/71 in every execution mode ✅.


🧠 Cache coherence, from first principles

Until now every hart shared one L1, so each line had exactly one copy and there was nothing to keep coherent. Now there is a choice:

coherence= Each core has What you learn
🏠 shared (default) the one shared L1 the baseline, unchanged
🌫️ none a private L1, no protocol 👻 a core keeps reading its own stale copy
🟦 msi private L1 + snooping bus Modified / Shared / Invalid
🟩 mesi private L1 + snooping bus ✨ Exclusive: the first write to an unshared line costs no bus request
🟧 moesi private L1 + snooping bus 🤝 Owned: dirty lines are shared, and the owner sends them cache to cache

🚌 The bus model:

  • Requests are BusRd, BusRdX, BusUpgr and BusWr.
  • Write-through stores go on the bus, and AMOs read the line for ownership.
  • An I-cache miss also snoops the D-caches, so code another core wrote is fetched current.
  • Snoops and cache-to-cache transfers cost cycles you can set: coherence.snoop_latency and coherence.c2c_latency.

🔬 Watch it happen

  • Settings → MULTI-CORE CACHES: protocol, snoop latency and cache-to-cache latency.
  • Cache → view: the selected core's L1. The dirty bit is replaced by the line state, colour-coded: 🟪 M · 🟧 O · 🟩 E · 🟦 S · ⬜ I.
  • Cache → coherence (new subtab): each core's bus traffic, the state of every recent line in every core, and a log of transitions:
  #instr  core  request   line        transitions
      23  c1    BusRd     0x00008000  c1 I→S, c0 M→S  written back first
      17  c0    BusRdX    0x00008010  c0 I→M, c1 E→I
      12  c1    BusRd     0x00008010  c1 I→E

⌨️ Or from the shell

raven run prog.s --cores 2 --coherence mesi --format csv

JSON, rstats and CSV now include each core's L1 and its bus counters:

  • requests sent;
  • invalidations received;
  • flushes;
  • cache-to-cache transfers;
  • coherence misses;
  • bus cycles.

🧪 Two experiments are included

🚩 harts/coherence_flag_handoff.s: core 0 raises a flag and core 1 waits for it.

🏠 shared 🌫️ none 🟦 MSI 🟩 MESI 🟧 MOESI
result ✅ 42 ♾️ never ends ✅ 42 ✅ 42 ✅ 42
cycles 772 — 1376 1376 1000
bus cycles — — 292 292 124

🏓 harts/coherence_false_sharing.s: two cores each bump their own counter 200 times.

MESI cycles bus requests invalidations
😵 counters on the same line 50 444 607 400
😌 counters 16 bytes apart 6 558 8 0

Neither core ever reads the other's counter, yet the line ping-pongs between the two caches on every write. That is false sharing. 📖 The full guide, with transition tables for all three protocols, is at docs/en/coherence.md · 🇧🇷 docs/pt-BR/coherence.md


🧬 One engine, four machines

Raven's backends are now trait-driven. Switch architectures with a on the Settings screen, or with --arch on the CLI:

--arch What it is
🐉 riscv32 RV32IMAF: VM/TLB, JIT, multicore, floating point
🧱 x86_64 an Intel-syntax integer/system core with its own split cache and pipeline view
🍼 sap the classic 8-bit Simple-As-Possible teaching CPU
🧸 toy16 a tiny 16-bit ISA, deliberately unlike RISC-V
  • 🗂️ The Run, Cache and Virtual Memory tabs are shared by every architecture, and each backend only shows what it actually models.
  • 📦 The simulator now lives in its own crate, raven-engine. Falcon runs a program headless for graders and tests:
let r = raven_engine::Falcon::new()
    .asm(".text\n li a0, 42\n li a7, 93\n ecall\n")
    .run()?;
assert_eq!(r.exit_code, Some(42));

🚦 Four pipeline models

mode= in the [pipeline] section selects one of four models: Serialized, Parallel UFs, Scoreboard and Tomasulo + ROB.

  • 🕸️ The out-of-order workbench is back, with hover and dependency wires.
  • 🗺️ The datapath is now a graph that each architecture declares.
  • 🩹 Before this release, a config asking for Scoreboard or Tomasulo silently ran the in-order datapath. It now builds the engine it asks for.

🧮 RV32F, by the book

  • 🚩 IEEE exception flags are raised.
  • ⚙️ fflags, frm and fcsr work.
  • 🧊 NaN results are canonical.
  • 🎯 fmin/fmax, fcvt.wu.s and fclass.s follow the spec.
  • 🧵 The fused ops round once. A new exact-integer FPU is shared by the sequential and pipelined modes. (#80)
  • 🎚️ Rounding modes rne/rtz/rdn/rup/rmm/dyn are honoured by every rounding instruction, written as an optional last operand. A reserved mode raises an illegal-instruction trap. (#87)

🛠️ Pipeline correctness

  • 🧾 Zicsr instructions really run in the pipeline. (#81)
  • 🔁 A result written to f0 is forwarded, and the out-of-order models see writes to f0. (#82, #86)
  • 🚚 flw/fsw go through the memory unit in every mode. (#83)
  • 🛫 mret, sret and sfence.vma run, and fetch resumes at mepc/sepc. (#84)
  • 🧽 fence.i makes stored code visible: dirty D-cache lines are written back before the I-cache is dropped. (#85)

✅ riscv-tests: 71 / 71, everywhere

The tests are built with the suite's official env/p, which needed HTIF tohost 📮 and a real mhartid 🪪. (#88)

suite seq Serialized Parallel UFs Scoreboard Tomasulo + ROB
rv32ui ✅ 42 ✅ 42 ✅ 42 ✅ 42 ✅ 42
rv32um ✅ 8 ✅ 8 ✅ 8 ✅ 8 ✅ 8
rv32ua ✅ 10 ✅ 10 ✅ 10 ✅ 10 ✅ 10
rv32uf ✅ 11 ✅ 11 ✅ 11 ✅ 11 ✅ 11

🤖 CI now runs cargo test on 🐧 Linux, 🪟 Windows and 🍎 macOS. An ELF fixture is committed, so a clean checkout passes. (#89)


📂 Real files for guest programs

openat, read, write, lseek, fstat, unlinkat, mkdirat… now work on real host files inside a chroot-like sandbox 🏖️. .. never escapes it. Linux ABI coverage is broader, and a valid Linux syscall never halts the run.


🎬 Polish

  • 🌅 A cinematic boot splash with a breathing PRESS ANY KEY.
  • 🖱️ Draggable scrollbars on register lists, TLB entries and cache snapshots.
  • 🧹 Fixes for screen inconsistencies across tabs.
  • 🗃️ One .rcfg ([sim]/[cache]/[pipeline]) and one .rstats file, and examples renamed from .fas to .s.

🐞 Issues closed

#80 RV32F flags/fcsr/NaN · #81 Zicsr in the pipeline · #82 f0 forwarding · #83 flw/fsw in Parallel UFs · #84 mret/sret/sfence.vma · #85 fence.i · #86 f0 in the OoO timing model · #87 rounding modes · #88 official env/p · #89 clean-checkout cargo test


⚠️ Upgrade notes

  • 🏠 Nothing changes unless you ask for it. coherence defaults to shared, and configs without the new keys load exactly as before.
  • 📝 raven export-config now writes coherence=, coherence.snoop_latency= and coherence.c2c_latency= in [cache].
  • 🔌 For raven-engine users: CacheLineView gained a state field, so add state: None where you build one. The new CacheHierarchy methods (core_count, core_cache, core_set, core_amat, coherence) all have defaults.
  • 🐢 The JIT runs whole blocks per turn, so for studying how cores interleave, keep the interpreter (--jit none, the default).

📦 Downloads

Platform File
🪟 Windows x86-64 raven-x86_64-pc-windows-msvc.zip
🐧 Linux x86-64 (glibc) raven-x86_64-unknown-linux-gnu.tar.gz
🐧 Linux x86-64 (static, musl) raven-x86_64-unknown-linux-musl.tar.gz
🍎 macOS Intel raven-x86_64-apple-darwin.tar.gz
🍏 macOS Apple Silicon raven-aarch64-apple-darwin.tar.gz

🔐 Checksums are in SHA256SUMS.txt.

Full changelog: v1.28.0...v1.29.0