Use SIMD vector reductions

Cargo.toml

@@ -18,7 +18,7 @@ bench = false
 simd-accel = ["faster"]
 
 [dependencies]
-faster = { version = "0.4.3", optional = true }
+faster = { git = "https://github.com/AdamNiederer/faster", branch="master", optional = true }
 
 [dev-dependencies]
 quickcheck = "0.6"

src/lib.rs

@@ -187,14 +187,17 @@ pub fn count(haystack: &[u8], needle: u8) -> usize {
             ret += (&haystack[i * u8s::WIDTH * 255..(i + 1) * u8s::WIDTH * 255])
                 .simd_iter()
                 .simd_reduce(u8s(0), u8s(needle.overflowing_add(1).0), |acc, v| {
-                    acc + (PackedEq::eq(&v, &u8s(needle)).be_u8s() & u8s(0x01))
-                }).scalar_reduce(0, |acc, s| acc + (s as usize));
+                    acc + v.eq_mask(u8s(needle)).be_u8s() & u8s(0x01)
+                }).sum_upcast() as usize;
         }
-        ret + (&haystack[haystack.len() - haystack.len() % (u8s::WIDTH * 255)..])
+
+        let final_chunk = (&haystack[haystack.len() - haystack.len() % (u8s::WIDTH * 255)..])
             .simd_iter()
             .simd_reduce(u8s(0), u8s(needle.overflowing_add(1).0), |acc, v| {
-                acc + (PackedEq::eq(&v, &u8s(needle)).be_u8s() & u8s(0x01))
-            }).scalar_reduce(0, |acc, s| acc + (s as usize))
+                acc + v.eq_mask(u8s(needle)).be_u8s() & u8s(0x01)
+            }).sum_upcast() as usize;
+
+        ret + final_chunk
     }
 }
 
@@ -276,7 +279,18 @@ pub fn num_chars(haystack: &[u8]) -> usize {
     num_chars_generic::<usize>(32, haystack)
 }
 
-/// f
+/// Count the number of UTF-8 encoded unicode codepoints in a slice of bytes, fast
+///
+/// This function is safe to use on any byte array, valid UTF-8 or not,
+/// but the output is only meaningful for well-formed UTF-8.
+///
+/// # Example
+///
+/// ```
+/// let swordfish = "メカジキ";
+/// let char_count = bytecount::num_chars(swordfish.as_bytes());
+/// assert_eq!(char_count, 4);
+/// ```
 #[cfg(feature = "simd-accel")]
 pub fn num_chars(haystack: &[u8]) -> usize {
     if haystack.len() < 100 {
@@ -288,14 +302,17 @@ pub fn num_chars(haystack: &[u8]) -> usize {
             ret += (&haystack[i * u8s::WIDTH * 255..(i + 1) * u8s::WIDTH * 255])
                 .simd_iter()
                 .simd_reduce(u8s(0), u8s(0), |acc, v| {
-                    acc + (PackedEq::eq(&(v & u8s(0xC0)), &u8s(0x80)).be_u8s() & u8s(0x01))
-                }).scalar_reduce(0, |acc, s| acc + (s as usize));
+                    acc + (v & u8s(0xC0)).eq_mask(u8s(0x80)).be_u8s() & u8s(0x01)
+                }).sum_upcast() as usize;
         }
-        haystack.len() - ret - (&haystack[haystack.len() - haystack.len() % (u8s::WIDTH * 255)..])
+
+        let final_chunk = (&haystack[haystack.len() - haystack.len() % (u8s::WIDTH * 255)..])
             .simd_iter()
             .simd_reduce(u8s(0), u8s(0), |acc, v| {
-                acc + (PackedEq::eq(&(v & u8s(0xC0)), &u8s(0x80)).be_u8s() & u8s(0x01))
-            }).scalar_reduce(0, |acc, s| acc + (s as usize))
+                acc + (v & u8s(0xC0)).eq_mask(u8s(0x80)).be_u8s() & u8s(0x01)
+            }).sum_upcast() as usize;
+
+        haystack.len() - ret - final_chunk
     }
 }