CPU work: SIMD Blake2b kernel (AVX-512 / AVX2 / NEON), ~5-7x faster than the current CPU path

.github/workflows/build.yml

@@ -8,7 +8,6 @@ jobs:
     steps:
       - name: Install OpenCL
         run: |
-          sudo add-apt-repository ppa:intel-opencl/intel-opencl
           sudo apt-get update
           sudo apt-get install ocl-icd-opencl-dev
       - uses: actions/checkout@aabbfeb2ce60b5bd82389903509092c4648a9713
@@ -22,7 +21,6 @@ jobs:
     steps:
       - name: Install OpenCL
         run: |
-          sudo add-apt-repository ppa:intel-opencl/intel-opencl
           sudo apt-get update
           sudo apt-get install ocl-icd-opencl-dev
       - uses: actions/checkout@aabbfeb2ce60b5bd82389903509092c4648a9713

.github/workflows/nano-work-simd.yml

@@ -0,0 +1,46 @@
+# CI for the nano-work-simd crate: build, test and lint on x86-64 (AVX2 and,
+# where the runner has it, AVX-512) and on Apple Silicon (NEON). The macOS job
+# is the CI job that compiles and executes the NEON kernel.
+name: nano-work-simd
+
+on:
+  push:
+    paths:
+      - "nano-work-simd/**"
+      - "src/**"
+      - "tests/**"
+      - "Cargo.toml"
+      - "Cargo.lock"
+      - ".github/workflows/nano-work-simd.yml"
+  pull_request:
+  workflow_dispatch:
+
+env:
+  CARGO_TERM_COLOR: always
+  RUST_BACKTRACE: 1
+
+jobs:
+  test:
+    name: test (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-14]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: clippy
+      - name: CPU features
+        shell: bash
+        run: |
+          uname -m
+          if [ "$(uname -s)" = "Linux" ]; then grep -m1 -o 'avx512f\|avx2' /proc/cpuinfo | sort -u; fi
+          if [ "$(uname -s)" = "Darwin" ]; then sysctl -n machdep.cpu.brand_string; fi
+      - name: Report kernels
+        run: cargo test --release -p nano-work-simd --lib report_kernels -- --nocapture
+      - name: Test
+        run: cargo test --release -p nano-work-simd
+      - name: Clippy
+        run: cargo clippy -p nano-work-simd --all-targets -- -D warnings

Cargo.toml

@@ -15,10 +15,13 @@ ocl = "0.19.4"
 serde_json = "1.0.87"
 hex = "0.4.3"
 rand = "0.8.5"
-blake2 = "0.10.4"
-digest = "0.10.5"
 byteorder = "1.4.3"
+nano-work-simd = { path = "nano-work-simd" }
 parking_lot = "0.12.1"
 chrono = "0.4.22"
 tokio = { version = "1.21.2", features = ["rt-multi-thread", "macros"] }
 rand_xorshift = "0.3.0"
+
+[workspace]
+members = ["nano-work-simd"]
+default-members = [".", "nano-work-simd"]

nano-work-simd/Cargo.toml

@@ -0,0 +1,20 @@
+[package]
+name = "nano-work-simd"
+version = "0.1.0"
+edition = "2021"
+rust-version = "1.89"
+license = "BSD-2-Clause"
+description = "Fast CPU proof-of-work generator for the Nano cryptocurrency (unrolled Blake2b-8 with AVX-512 / AVX2 / NEON / scalar kernels)"
+repository = "https://github.com/nanocurrency/nano-work-server"
+keywords = ["nano", "blake2b", "proof-of-work", "simd"]
+categories = ["cryptography", "algorithms"]
+
+[lib]
+name = "nano_work_simd"
+path = "src/lib.rs"
+
+[dependencies]
+# none: the crate is std-only
+
+[dev-dependencies]
+blake2 = "0.10"

nano-work-simd/README.md

@@ -0,0 +1,29 @@
+# nano-work-simd
+
+The CPU work back end of nano-work-server: Nano proof-of-work
+(`u64::from_le_bytes(blake2b_8(work_le || hash)) >= threshold`) with a Blake2b
+compression specialised for the fixed 40-byte message, run on 8 (AVX-512),
+4 (AVX2), 2 (NEON) or 1 (scalar) nonces per instruction stream. The kernel is
+picked at runtime from the CPU's feature flags (`Kernel::best()`), or forced by
+the server's `--cpu-kernel` option.
+
+Public API: `generate`, `generate_with`, `validate`, `difficulty`,
+`hash_rate_benchmark`, `expected_hashes`, `resolve_hit`, `Kernel`
+(`search`, `difficulties`, `best`, `available`, `from_name`), `Precomputed`,
+and the `SEND_THRESHOLD` / `RECEIVE_THRESHOLD` / `LEGACY_THRESHOLD` / `BATCH`
+constants. The server uses `Kernel::search` over `BATCH`-sized nonce ranges,
+`resolve_hit` to pick and re-verify the winning lane, and `validate` for
+`work_validate`.
+
+Tests (`cargo test --release -p nano-work-simd` from the server root) cross-check
+every available kernel against the `blake2` crate on 10 000 random (hash, nonce)
+pairs, check nonce wrap-around at `u64::MAX`, check that `search` finds exactly
+the hit set of a brute-force scan, that generated work validates for every
+kernel and thread count, and the documented Nano example. `Precomputed`'s words
+and the lane-level kernels are crate-private; `unsafe` is confined to the
+`#[target_feature]` entry points in `src/x86.rs` / `src/neon.rs`.
+
+This is a vendored copy of the standalone `nano-work-simd` crate, which also
+carries a CLI (`info`, `generate`, `validate`, `bench`, `timing`) and a
+`hashlib`-based Python cross-check. No dependencies; `blake2` is a
+dev-dependency for the tests only. MSRV 1.89. License: BSD-2-Clause.