Issue 210: Fast path hardening take 2 (#215)

* Reverts the basic fast path for u64 to be two folded multiplies.

Signed-off-by: Tom Kaitchuck <[email protected]>

Author: tkaitchuck

Cargo.toml

@@ -100,6 +100,7 @@ rand = "0.8.5"
 pcg-mwc = "0.2.1"
 serde_json = "1.0.59"
 hashbrown = "0.14.3"
+smallvec = "1.13.1"
 
 [package.metadata.docs.rs]
 rustc-args = ["-C", "target-feature=+aes"]

compare/Cargo.toml

@@ -29,6 +29,9 @@ codegen-units = 1
 
 [dependencies]
 ahash = { path = "../", default-features = false }
+pcg-mwc = "0.2.1"
+rand = "0.8.5"
+rand_core = "0.6.4"
 
 [dev-dependencies]
 criterion = "0.3.3"

compare/src/main.rs

@@ -0,0 +1,32 @@
+use std::io::Error;
+use std::fs::File;
+use std::io::Write;
+use pcg_mwc::Mwc256XXA64;
+use ahash::RandomState;
+use std::io::BufWriter;
+use std::path::Path;
+use rand_core::SeedableRng;
+use rand::Rng;
+use std::time::Instant;
+
+
+fn main() -> Result<(), Error> {
+    let mut r = Mwc256XXA64::seed_from_u64(0xe786_c22b_119c_1479);
+
+    let path = Path::new("hash_output");
+
+    let mut file = BufWriter::new(File::create(path)?);
+    let hasher = RandomState::with_seeds(r.gen(), r.gen(), r.gen(), r.gen());
+    let start = Instant::now();
+    let mut sum: u64 = 0;
+    for i in 0..i32::MAX {
+        let value = hasher.hash_one(i as u64);
+        sum = sum.wrapping_add(value);
+        let value: [u8; 8] = value.to_ne_bytes();
+        file.write_all(&value)?;
+    }
+    let elapsed = start.elapsed();
+    println!("Sum {} Elapsed time: {}", sum, elapsed.as_millis());
+    file.flush()?;
+    Ok(())
+}

src/aes_hash.rs

@@ -225,8 +225,7 @@ pub(crate) struct AHasherU64 {
 impl Hasher for AHasherU64 {
     #[inline]
     fn finish(&self) -> u64 {
-        let rot = (self.pad & 63) as u32;
-        self.buffer.rotate_left(rot)
+        folded_multiply(self.buffer, self.pad)
     }
 
     #[inline]
@@ -252,7 +251,6 @@ impl Hasher for AHasherU64 {
     #[inline]
     fn write_u64(&mut self, i: u64) {
         self.buffer = folded_multiply(i ^ self.buffer, MULTIPLE);
-        self.pad = self.pad.wrapping_add(i);
     }
 
     #[inline]

src/fallback_hash.rs

@@ -56,8 +56,8 @@ impl AHasher {
     #[allow(dead_code)] // Is not called if non-fallback hash is used.
     pub(crate) fn from_random_state(rand_state: &RandomState) -> AHasher {
         AHasher {
-            buffer: rand_state.k0,
-            pad: rand_state.k1,
+            buffer: rand_state.k1,
+            pad: rand_state.k0,
             extra_keys: [rand_state.k2, rand_state.k3],
         }
     }
@@ -117,7 +117,7 @@ impl AHasher {
     #[inline]
     #[cfg(feature = "specialize")]
     fn short_finish(&self) -> u64 {
-        self.buffer.wrapping_add(self.pad)
+        folded_multiply(self.buffer, self.pad)
     }
 }
 
@@ -210,8 +210,8 @@ pub(crate) struct AHasherU64 {
 impl Hasher for AHasherU64 {
     #[inline]
     fn finish(&self) -> u64 {
-        let rot = (self.pad & 63) as u32;
-        self.buffer.rotate_left(rot)
+        folded_multiply(self.buffer, self.pad)
+        //self.buffer
     }
 
     #[inline]
@@ -237,7 +237,6 @@ impl Hasher for AHasherU64 {
     #[inline]
     fn write_u64(&mut self, i: u64) {
         self.buffer = folded_multiply(i ^ self.buffer, MULTIPLE);
-        self.pad = self.pad.wrapping_add(i);
     }
 
     #[inline]

src/hash_quality_test.rs

@@ -338,22 +338,24 @@ fn test_length_extension<T: Hasher>(hasher: impl Fn(u128, u128) -> T) {
 }
 
 fn test_sparse<T: Hasher>(hasher: impl Fn() -> T) {
+    use smallvec::SmallVec;
+
     let mut buf = [0u8; 256];
     let mut hashes = HashMap::new();
-    for idx_1 in 0..256 {
-        for idx_2 in idx_1 + 1..256 {
+    for idx_1 in 0..255_u8 {
+        for idx_2 in idx_1 + 1..=255_u8 {
             for value_1 in [1, 2, 4, 8, 16, 32, 64, 128] {
                 for value_2 in [
                     1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 15, 16, 17, 18, 20, 24, 31, 32, 33, 48, 64, 96, 127, 128, 129,
                     192, 254, 255,
                 ] {
-                    buf[idx_1] = value_1;
-                    buf[idx_2] = value_2;
+                    buf[idx_1 as usize] = value_1;
+                    buf[idx_2 as usize] = value_2;
                     let hash_value = hash_with(&buf, &mut hasher());
-                    let keys = hashes.entry(hash_value).or_insert(Vec::new());
-                    keys.push((idx_1, value_1, idx_2, value_2));
-                    buf[idx_1] = 0;
-                    buf[idx_2] = 0;
+                    let keys = hashes.entry(hash_value).or_insert(SmallVec::<[[u8; 4]; 1]>::new());
+                    keys.push([idx_1, value_1, idx_2, value_2]);
+                    buf[idx_1 as usize] = 0;
+                    buf[idx_2 as usize] = 0;
                 }
             }
         }

src/random_state.rs

@@ -470,8 +470,8 @@ impl BuildHasherExt for RandomState {
     #[inline]
     fn hash_as_u64<T: Hash + ?Sized>(&self, value: &T) -> u64 {
         let mut hasher = AHasherU64 {
-            buffer: self.k0,
-            pad: self.k1,
+            buffer: self.k1,
+            pad: self.k0,
         };
         value.hash(&mut hasher);
         hasher.finish()

tests/bench.rs

@@ -75,7 +75,7 @@ fn gen_strings() -> Vec<String> {
 macro_rules! bench_inputs {
     ($group:ident, $hash:ident) => {
         // Number of iterations per batch should be high enough to hide timing overhead.
-        let size = BatchSize::NumIterations(2_000);
+        let size = BatchSize::NumIterations(50_000);
 
         let mut rng = rand::thread_rng();
         $group.bench_function("u8", |b| b.iter_batched(|| rng.gen::<u8>(), |v| $hash(&v), size));

tests/map_tests.rs

@@ -238,7 +238,7 @@ fn test_byte_dist() {
         let mut table: [bool; 256 * 8] = [false; 256 * 8];
         let hasher = RandomState::with_seeds(r.gen(), r.gen(), r.gen(), r.gen());
         for i in 0..128 {
-            let mut keys: [u8; 8] = hasher.hash_one(i as u64).to_ne_bytes();
+            let mut keys: [u8; 8] = hasher.hash_one((i as u64) << 30).to_ne_bytes();
             //let mut keys = r.next_u64().to_ne_bytes(); //This is a control to test assert sensitivity.
             for idx in 0..8 {
                 while table[idx * 256 + keys[idx] as usize] {