Stable simd

.github/workflows/test.yml

@@ -35,8 +35,17 @@ jobs:
           toolchain: ${{ matrix.rust }}
           components: rustfmt, clippy
 
+      # Builds use `-C target-cpu=native`, so cached artifacts (including proc macros, which run in
+      # the compiler) only work on runners with the same target features
+      - name: Identify native target features
+        id: native
+        run: |
+          echo "features=$(rustc --print cfg -C target-cpu=native | grep '^target_feature=' | sha256sum | cut -c1-12)" >> "$GITHUB_OUTPUT"
+
       - name: Caching
         uses: Swatinem/rust-cache@v2
+        with:
+          key: native-${{ steps.native.outputs.features }}
 
       - name: Check
         # clippy will also do a build check
@@ -48,10 +57,6 @@ jobs:
           cargo clippy -p roaring --all-targets --no-default-features -- -D warnings
           cargo clippy -p roaring --all-targets --features serde -- -Dclippy::todo -D warnings
 
-      - name: Check SIMD
-        if: matrix.rust == 'nightly'
-        run: cargo clippy -p roaring --all-targets --all-features -- -D warnings
-
       - name: Check formatting
         run: cargo fmt --all -- --check
 
@@ -68,9 +73,6 @@ jobs:
         features:
           - default
           - no-std
-        include:
-          - rust: nightly
-            features: simd
     env:
       RUSTFLAGS: "-C target-cpu=native -C opt-level=3"
       ROARINGRS_BENCH_OFFLINE: "true"
@@ -85,8 +87,17 @@ jobs:
           repository: "RoaringBitmap/real-roaring-datasets"
           path: "benchmarks/real-roaring-datasets"
 
+      # Builds use `-C target-cpu=native`, so cached artifacts (including proc macros, which run in
+      # the compiler) only work on runners with the same target features
+      - name: Identify native target features
+        id: native
+        run: |
+          echo "features=$(rustc --print cfg -C target-cpu=native | grep '^target_feature=' | sha256sum | cut -c1-12)" >> "$GITHUB_OUTPUT"
+
       - name: Caching
         uses: Swatinem/rust-cache@v2
+        with:
+          key: native-${{ steps.native.outputs.features }}
 
       - name: Initialize rust toolchain
         uses: dtolnay/rust-toolchain@master
@@ -105,10 +116,6 @@ jobs:
         if: matrix.features == 'no-std'
         run: cargo test -p roaring --no-default-features
 
-      - name: SIMD test
-        if: matrix.rust == 'nightly' && matrix.features == 'simd'
-        run: cargo +nightly test -p roaring --features simd
-
   miri:
     runs-on: ubuntu-latest
     needs: build
@@ -126,8 +133,17 @@ jobs:
           toolchain: nightly
           components: miri
 
+      # Builds use `-C target-cpu=native`, so cached artifacts (including proc macros, which run in
+      # the compiler) only work on runners with the same target features
+      - name: Identify native target features
+        id: native
+        run: |
+          echo "features=$(rustc --print cfg -C target-cpu=native | grep '^target_feature=' | sha256sum | cut -c1-12)" >> "$GITHUB_OUTPUT"
+
       - name: Caching
         uses: Swatinem/rust-cache@v2
+        with:
+          key: native-${{ steps.native.outputs.features }}
 
       - name: Setup miri
         run: cargo miri setup
@@ -150,8 +166,17 @@ jobs:
         with:
           toolchain: nightly
 
+      # Builds use `-C target-cpu=native`, so cached artifacts (including proc macros, which run in
+      # the compiler) only work on runners with the same target features
+      - name: Identify native target features
+        id: native
+        run: |
+          echo "features=$(rustc --print cfg -C target-cpu=native | grep '^target_feature=' | sha256sum | cut -c1-12)" >> "$GITHUB_OUTPUT"
+
       - name: Caching
         uses: Swatinem/rust-cache@v2
+        with:
+          key: native-${{ steps.native.outputs.features }}
 
       - name: Install cargo fuzz
         run: cargo install cargo-fuzz
@@ -165,7 +190,4 @@ jobs:
             fuzz/corpus
 
       - name: Run Fuzzer vs croaring for 15 minutes
-        run: cargo fuzz run against_croaring -s none -- -timeout=5 -max_total_time=900
-
-      - name: Run Fuzzer (with simd) vs croaring for 15 minutes
-        run: cargo fuzz run --features=simd against_croaring -- -timeout=5 -max_total_time=900
+        run: cargo fuzz run against_croaring -- -timeout=5 -max_total_time=900

Cargo.toml

@@ -8,6 +8,8 @@ roaring = { path = "roaring" }
 bytemuck = "1.25.2"
 byteorder = "1.5.0"
 criterion = "0.5"
+fearless_simd = { version = "1.0.0", default-features = false }
+fearless_simd_macros = "0.1.0"
 git2 = { version = "0.21", default-features = false }
 indicatif = "0.17"
 itertools = "0.15"

README.md

@@ -68,7 +68,10 @@ https://github.com/RoaringBitmap/roaring-rs/actions/workflows/test.yml/badge.svg
 [real-world datasets]: https://github.com/RoaringBitmap/real-roaring-datasets
 [on the user guide]: https://bheisler.github.io/criterion.rs/book/user_guide/user_guide.html
 
-## Experimental features
+## SIMD
 
-The `simd` feature is in active development. It has not been tested. If you would like to build with `simd` note that
-`std::simd` is only available in Rust nightly.
+Some operations are accelerated with SIMD, implemented with the [fearless_simd] crate. When the `std` feature is
+enabled, the best available SIMD instruction set is detected at runtime on x86 and x86-64; otherwise, it is determined
+by the target features enabled at compile time (e.g. with `-C target-cpu=native`).
+
+[fearless_simd]: https://docs.rs/fearless_simd

benchmarks/Cargo.toml

@@ -17,9 +17,6 @@ indicatif = { workspace = true }
 itertools = { workspace = true }
 zip = { workspace = true, default-features = false, features = ["deflate"] }
 
-[features]
-simd = ["roaring/simd"]
-
 [[bench]]
 name = "lib"
 harness = false

fuzz/Cargo.toml

@@ -12,9 +12,6 @@ libfuzzer-sys = { version = "0.4.9", features = ["arbitrary-derive"] }
 roaring = { path = "../roaring" }
 croaring = "2.0"
 
-[features]
-simd = ["roaring/simd"]
-
 [[bin]]
 name = "against_croaring"
 path = "fuzz_targets/against_croaring.rs"

roaring/Cargo.toml

@@ -20,12 +20,14 @@ license = "MIT OR Apache-2.0"
 bytemuck = { workspace = true, optional = true }
 byteorder = { workspace = true, optional = true }
 serde = { workspace = true, optional = true }
+# At least one of std and libm is required; std overrides libm.
+fearless_simd = { workspace = true, features = ["libm"] }
+fearless_simd_macros = { workspace = true }
 
 [features]
 default = ["std"]
 serde = ["dep:serde", "std"]
-simd = []
-std = ["dep:bytemuck", "dep:byteorder"]
+std = ["dep:bytemuck", "dep:byteorder", "fearless_simd/std"]
 
 [dev-dependencies]
 proptest = { workspace = true }

roaring/src/bitmap/store/array_store/mod.rs

@@ -256,10 +256,7 @@ impl ArrayStore {
 
     pub fn intersection_len(&self, other: &Self) -> u64 {
         let mut visitor = CardinalityCounter::new();
-        #[cfg(feature = "simd")]
         vector::and(self.as_slice(), other.as_slice(), &mut visitor);
-        #[cfg(not(feature = "simd"))]
-        scalar::and(self.as_slice(), other.as_slice(), &mut visitor);
         visitor.into_inner()
     }
 
@@ -418,10 +415,7 @@ impl BitOr<Self> for &ArrayStore {
         #[allow(clippy::suspicious_arithmetic_impl)]
         let capacity = self.vec.len() + rhs.vec.len();
         let mut visitor = VecWriter::new(capacity);
-        #[cfg(feature = "simd")]
         vector::or(self.as_slice(), rhs.as_slice(), &mut visitor);
-        #[cfg(not(feature = "simd"))]
-        scalar::or(self.as_slice(), rhs.as_slice(), &mut visitor);
         ArrayStore::from_vec_unchecked(visitor.into_inner())
     }
 }
@@ -431,31 +425,16 @@ impl BitAnd<Self> for &ArrayStore {
 
     fn bitand(self, rhs: Self) -> Self::Output {
         let mut visitor = VecWriter::new(self.vec.len().min(rhs.vec.len()));
-        #[cfg(feature = "simd")]
         vector::and(self.as_slice(), rhs.as_slice(), &mut visitor);
-        #[cfg(not(feature = "simd"))]
-        scalar::and(self.as_slice(), rhs.as_slice(), &mut visitor);
         ArrayStore::from_vec_unchecked(visitor.into_inner())
     }
 }
 
 impl BitAndAssign<&Self> for ArrayStore {
-    #[allow(clippy::suspicious_op_assign_impl)]
     fn bitand_assign(&mut self, rhs: &Self) {
-        #[cfg(feature = "simd")]
-        {
-            let mut visitor = VecWriter::new(self.vec.len().min(rhs.vec.len()));
-            vector::and(self.as_slice(), rhs.as_slice(), &mut visitor);
-            self.vec = visitor.into_inner()
-        }
-        #[cfg(not(feature = "simd"))]
-        {
-            let mut i = 0;
-            self.retain(|x| {
-                i += rhs.iter().skip(i).position(|y| *y >= x).unwrap_or(rhs.vec.len());
-                rhs.vec.get(i).is_some_and(|y| x == *y)
-            });
-        }
+        let mut visitor = VecWriter::new(self.vec.len().min(rhs.vec.len()));
+        vector::and(self.as_slice(), rhs.as_slice(), &mut visitor);
+        self.vec = visitor.into_inner()
     }
 }
 
@@ -470,31 +449,16 @@ impl Sub<Self> for &ArrayStore {
 
     fn sub(self, rhs: Self) -> Self::Output {
         let mut visitor = VecWriter::new(self.vec.len());
-        #[cfg(feature = "simd")]
         vector::sub(self.as_slice(), rhs.as_slice(), &mut visitor);
-        #[cfg(not(feature = "simd"))]
-        scalar::sub(self.as_slice(), rhs.as_slice(), &mut visitor);
         ArrayStore::from_vec_unchecked(visitor.into_inner())
     }
 }
 
 impl SubAssign<&Self> for ArrayStore {
-    #[allow(clippy::suspicious_op_assign_impl)]
     fn sub_assign(&mut self, rhs: &Self) {
-        #[cfg(feature = "simd")]
-        {
-            let mut visitor = VecWriter::new(self.vec.len().min(rhs.vec.len()));
-            vector::sub(self.as_slice(), rhs.as_slice(), &mut visitor);
-            self.vec = visitor.into_inner()
-        }
-        #[cfg(not(feature = "simd"))]
-        {
-            let mut i = 0;
-            self.retain(|x| {
-                i += rhs.iter().skip(i).position(|y| *y >= x).unwrap_or(rhs.vec.len());
-                rhs.vec.get(i).is_none_or(|y| x != *y)
-            });
-        }
+        let mut visitor = VecWriter::new(self.vec.len().min(rhs.vec.len()));
+        vector::sub(self.as_slice(), rhs.as_slice(), &mut visitor);
+        self.vec = visitor.into_inner()
     }
 }
 
@@ -511,10 +475,7 @@ impl BitXor<Self> for &ArrayStore {
         #[allow(clippy::suspicious_arithmetic_impl)]
         let capacity = self.vec.len() + rhs.vec.len();
         let mut visitor = VecWriter::new(capacity);
-        #[cfg(feature = "simd")]
         vector::xor(self.as_slice(), rhs.as_slice(), &mut visitor);
-        #[cfg(not(feature = "simd"))]
-        scalar::xor(self.as_slice(), rhs.as_slice(), &mut visitor);
         ArrayStore::from_vec_unchecked(visitor.into_inner())
     }
 }