chacha20: add an AVX-512 backend

.github/workflows/chacha20.yml

@@ -20,6 +20,9 @@ defaults:
 env:
   CARGO_INCREMENTAL: 0
   RUSTFLAGS: "-Dwarnings"
+  # NOTE: The mirror number changes with each version so keep these in sync
+  SDE_FULL_VERSION_MIRROR: "859732"
+  SDE_FULL_VERSION: "9.58.0-2025-06-16"
 
 jobs:
   build:
@@ -79,6 +82,43 @@ jobs:
       - run: cargo check --target ${{ matrix.target }} --all-features
       - run: cargo hack test --feature-powerset --target ${{ matrix.target }}
 
+  # Tests for the AVX-512 backend
+  avx512:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        include:
+          - target: x86_64-unknown-linux-gnu
+            rust: stable
+            RUSTFLAGS: "-Dwarnings --cfg chacha20_avx512"
+    env:
+      CARGO_INCREMENTAL: 0
+      RUSTFLAGS: ${{ matrix.RUSTFLAGS }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install Intel SDE
+        run: |
+          curl -JLO "https://downloadmirror.intel.com/${{ env.SDE_FULL_VERSION_MIRROR }}/sde-external-${{ env.SDE_FULL_VERSION }}-lin.tar.xz"
+          tar xvf sde-external-${{ env.SDE_FULL_VERSION }}-lin.tar.xz -C /opt
+          echo "/opt/sde-external-${{ env.SDE_FULL_VERSION }}-lin" >> $GITHUB_PATH
+      - uses: RustCrypto/actions/cargo-cache@master
+      - uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: ${{ matrix.rust }}
+          targets: ${{ matrix.target }}
+      # NOTE: Write a `.cargo/config.toml` to configure the target for AVX-512
+      # NOTE: We use intel-sde as the runner since not all GitHub CI hosts support AVX512
+      - name: write .cargo/config.toml
+        shell: bash
+        run: |
+          cd ../chacha20/..
+          mkdir -p .cargo
+          echo '[target.${{ matrix.target }}]'  > .cargo/config.toml
+          echo 'runner = "sde64 -future --"'   >> .cargo/config.toml
+      - run: ${{ matrix.deps }}
+      - run: cargo test --target ${{ matrix.target }}
+      - run: cargo test --target ${{ matrix.target }} --all-features
+
   # Tests for the AVX2 backend
   avx2:
     runs-on: ubuntu-latest

chacha20/Cargo.toml

@@ -34,6 +34,7 @@ cipher = { version = "0.5.0-rc.1", features = ["dev"] }
 hex-literal = "1"
 proptest = "1"
 rand_chacha = "0.9"
+hex = "0.4"
 
 [features]
 default = ["cipher"]
@@ -48,9 +49,11 @@ rustdoc-args = ["--cfg", "docsrs"]
 [lints.rust.unexpected_cfgs]
 level = "warn"
 check-cfg = [
+    'cfg(chacha20_avx512)',
     'cfg(chacha20_force_soft)',
     'cfg(chacha20_force_sse2)',
     'cfg(chacha20_force_avx2)',
+    'cfg(chacha20_force_avx512)',
 ]
 
 [lints.clippy]

chacha20/README.md

@@ -32,6 +32,7 @@ work on stable Rust with the following `RUSTFLAGS`:
 - `x86` / `x86_64`
   - `avx2`: (~1.4cpb) `-Ctarget-cpu=haswell -Ctarget-feature=+avx2`
   - `sse2`: (~1.6cpb) `-Ctarget-feature=+sse2` (on by default on x86 CPUs)
+  - `avx512`: `-Ctarget-feature=+avx512f,+avx512vl --cfg chacha20_avx512` requires Rust 1.89+
 - `aarch64`
   - `neon` (~2-3x faster than `soft`) requires Rust 1.61+ and the `neon` feature enabled
 - Portable

chacha20/src/backends.rs

@@ -7,12 +7,19 @@ cfg_if! {
         pub(crate) mod soft;
     } else if #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] {
         cfg_if! {
-            if #[cfg(chacha20_force_avx2)] {
+            if #[cfg(all(chacha20_avx512, chacha20_force_avx512))] {
+                pub(crate) mod avx512;
+                // AVX-2 backend needed for RNG if enabled
+                #[cfg(feature = "rng")]
+                pub(crate) mod avx2;
+            } else if #[cfg(chacha20_force_avx2)] {
                 pub(crate) mod avx2;
             } else if #[cfg(chacha20_force_sse2)] {
                 pub(crate) mod sse2;
             } else {
                 pub(crate) mod soft;
+                #[cfg(chacha20_avx512)]
+                pub(crate) mod avx512;
                 pub(crate) mod avx2;
                 pub(crate) mod sse2;
             }

chacha20/src/backends/avx2.rs

@@ -27,6 +27,7 @@ const N: usize = PAR_BLOCKS / 2;
 #[inline]
 #[target_feature(enable = "avx2")]
 #[cfg(feature = "cipher")]
+#[cfg_attr(chacha20_force_avx512, expect(unused))]
 pub(crate) unsafe fn inner<R, F, V>(state: &mut [u32; STATE_WORDS], f: F)
 where
     R: Rounds,