Improved performance of comparison with SIMD feature flag (2x-3.5x)

benches/comparison_kernels.rs

@@ -32,7 +32,7 @@ where
 
 fn bench_op_scalar<T>(arr_a: &PrimitiveArray<T>, value_b: T, op: Operator)
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
     primitive_compare_scalar(
         criterion::black_box(arr_a),

src/compute/comparison/mod.rs

@@ -30,6 +30,9 @@ mod boolean;
 mod primitive;
 mod utf8;
 
+mod simd;
+pub use simd::{Simd8, Simd8Lanes};
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum Operator {
     Lt,

src/compute/comparison/primitive.rs

@@ -23,45 +23,33 @@ use crate::{
     error::{ArrowError, Result},
 };
 
+use super::simd::{Simd8, Simd8Lanes};
 use super::{super::utils::combine_validities, Operator};
 
 pub(crate) fn compare_values_op<T, F>(lhs: &[T], rhs: &[T], op: F) -> MutableBitmap
 where
-    T: NativeType,
-    F: Fn(T, T) -> bool,
+    T: NativeType + Simd8,
+    F: Fn(T::Simd, T::Simd) -> u8,
 {
     assert_eq!(lhs.len(), rhs.len());
-    let mut values = MutableBuffer::from_len_zeroed((lhs.len() + 7) / 8);
 
     let lhs_chunks_iter = lhs.chunks_exact(8);
     let lhs_remainder = lhs_chunks_iter.remainder();
     let rhs_chunks_iter = rhs.chunks_exact(8);
     let rhs_remainder = rhs_chunks_iter.remainder();
 
-    let chunks = lhs.len() / 8;
-
-    values[..chunks]
-        .iter_mut()
-        .zip(lhs_chunks_iter)
-        .zip(rhs_chunks_iter)
-        .for_each(|((byte, lhs), rhs)| {
-            lhs.iter()
-                .zip(rhs.iter())
-                .enumerate()
-                .for_each(|(i, (&lhs, &rhs))| {
-                    *byte |= if op(lhs, rhs) { 1 << i } else { 0 };
-                });
-        });
+    let mut values = MutableBuffer::with_capacity((lhs.len() + 7) / 8);
+    let iterator = lhs_chunks_iter.zip(rhs_chunks_iter).map(|(lhs, rhs)| {
+        let lhs = T::Simd::from_chunk(lhs);
+        let rhs = T::Simd::from_chunk(rhs);
+        op(lhs, rhs)
+    });
+    values.extend_from_trusted_len_iter(iterator);
 
     if !lhs_remainder.is_empty() {
-        let last = &mut values[chunks];
-        lhs_remainder
-            .iter()
-            .zip(rhs_remainder.iter())
-            .enumerate()
-            .for_each(|(i, (&lhs, &rhs))| {
-                *last |= if op(lhs, rhs) { 1 << i } else { 0 };
-            });
+        let lhs = T::Simd::from_incomplete_chunk(lhs_remainder, T::default());
+        let rhs = T::Simd::from_incomplete_chunk(rhs_remainder, T::default());
+        values.push(op(lhs, rhs))
     };
     MutableBitmap::from_buffer(values, lhs.len())
 }
@@ -70,8 +58,8 @@ where
 /// comparison function.
 fn compare_op<T, F>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>, op: F) -> Result<BooleanArray>
 where
-    T: NativeType,
-    F: Fn(T, T) -> bool,
+    T: NativeType + Simd8,
+    F: Fn(T::Simd, T::Simd) -> u8,
 {
     if lhs.len() != rhs.len() {
         return Err(ArrowError::InvalidArgumentError(
@@ -90,31 +78,25 @@ where
 /// a specified comparison function.
 pub fn compare_op_scalar<T, F>(lhs: &PrimitiveArray<T>, rhs: T, op: F) -> Result<BooleanArray>
 where
-    T: NativeType,
-    F: Fn(T, T) -> bool,
+    T: NativeType + Simd8,
+    F: Fn(T::Simd, T::Simd) -> u8,
 {
     let validity = lhs.validity().clone();
-
-    let mut values = MutableBuffer::from_len_zeroed((lhs.len() + 7) / 8);
+    let rhs = T::Simd::from_chunk(&[rhs; 8]);
 
     let lhs_chunks_iter = lhs.values().chunks_exact(8);
     let lhs_remainder = lhs_chunks_iter.remainder();
-    let chunks = lhs.len() / 8;
 
-    values[..chunks]
-        .iter_mut()
-        .zip(lhs_chunks_iter)
-        .for_each(|(byte, chunk)| {
-            chunk.iter().enumerate().for_each(|(i, &c_i)| {
-                *byte |= if op(c_i, rhs) { 1 << i } else { 0 };
-            });
-        });
+    let mut values = MutableBuffer::with_capacity((lhs.len() + 7) / 8);
+    let iterator = lhs_chunks_iter.map(|lhs| {
+        let lhs = T::Simd::from_chunk(lhs);
+        op(lhs, rhs)
+    });
+    values.extend_from_trusted_len_iter(iterator);
 
     if !lhs_remainder.is_empty() {
-        let last = &mut values[chunks];
-        lhs_remainder.iter().enumerate().for_each(|(i, &lhs)| {
-            *last |= if op(lhs, rhs) { 1 << i } else { 0 };
-        });
+        let lhs = T::Simd::from_incomplete_chunk(lhs_remainder, T::default());
+        values.push(op(lhs, rhs))
     };
 
     Ok(BooleanArray::from_data(
@@ -126,105 +108,105 @@ where
 /// Perform `lhs == rhs` operation on two arrays.
 pub fn eq<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a == b)
+    compare_op(lhs, rhs, |a, b| a.eq(b))
 }
 
 /// Perform `left == right` operation on an array and a scalar value.
 pub fn eq_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a == b)
+    compare_op_scalar(lhs, rhs, |a, b| a.eq(b))
 }
 
 /// Perform `left != right` operation on two arrays.
 pub fn neq<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a != b)
+    compare_op(lhs, rhs, |a, b| a.neq(b))
 }
 
 /// Perform `left != right` operation on an array and a scalar value.
 pub fn neq_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a != b)
+    compare_op_scalar(lhs, rhs, |a, b| a.neq(b))
 }
 
 /// Perform `left < right` operation on two arrays.
 pub fn lt<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a < b)
+    compare_op(lhs, rhs, |a, b| a.lt(b))
 }
 
 /// Perform `left < right` operation on an array and a scalar value.
 pub fn lt_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a < b)
+    compare_op_scalar(lhs, rhs, |a, b| a.lt(b))
 }
 
 /// Perform `left <= right` operation on two arrays.
 pub fn lt_eq<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a <= b)
+    compare_op(lhs, rhs, |a, b| a.lt_eq(b))
 }
 
 /// Perform `left <= right` operation on an array and a scalar value.
 /// Null values are less than non-null values.
 pub fn lt_eq_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a <= b)
+    compare_op_scalar(lhs, rhs, |a, b| a.lt_eq(b))
 }
 
 /// Perform `left > right` operation on two arrays. Non-null values are greater than null
 /// values.
 pub fn gt<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a > b)
+    compare_op(lhs, rhs, |a, b| a.gt(b))
 }
 
 /// Perform `left > right` operation on an array and a scalar value.
 /// Non-null values are greater than null values.
 pub fn gt_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a > b)
+    compare_op_scalar(lhs, rhs, |a, b| a.gt(b))
 }
 
 /// Perform `left >= right` operation on two arrays. Non-null values are greater than null
 /// values.
 pub fn gt_eq<T>(lhs: &PrimitiveArray<T>, rhs: &PrimitiveArray<T>) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op(lhs, rhs, |a, b| a >= b)
+    compare_op(lhs, rhs, |a, b| a.gt_eq(b))
 }
 
 /// Perform `left >= right` operation on an array and a scalar value.
 /// Non-null values are greater than null values.
 pub fn gt_eq_scalar<T>(lhs: &PrimitiveArray<T>, rhs: T) -> Result<BooleanArray>
 where
-    T: NativeType + std::cmp::PartialOrd,
+    T: NativeType + Simd8,
 {
-    compare_op_scalar(lhs, rhs, |a, b| a >= b)
+    compare_op_scalar(lhs, rhs, |a, b| a.gt_eq(b))
 }
 
-pub fn compare<T: NativeType + std::cmp::PartialOrd>(
+pub fn compare<T: NativeType + Simd8>(
     lhs: &PrimitiveArray<T>,
     rhs: &PrimitiveArray<T>,
     op: Operator,
@@ -239,7 +221,7 @@ pub fn compare<T: NativeType + std::cmp::PartialOrd>(
     }
 }
 
-pub fn compare_scalar<T: NativeType + std::cmp::PartialOrd>(
+pub fn compare_scalar<T: NativeType + Simd8>(
     lhs: &PrimitiveArray<T>,
     rhs: T,
     op: Operator,

src/compute/comparison/simd/mod.rs

@@ -0,0 +1,91 @@
+use crate::types::NativeType;
+
+/// [`NativeType`] that supports a representation of 8 lanes
+pub trait Simd8: NativeType {
+    type Simd: Simd8Lanes<Self>;
+}
+
+pub trait Simd8Lanes<T>: Copy {
+    fn from_chunk(v: &[T]) -> Self;
+    fn from_incomplete_chunk(v: &[T], remaining: T) -> Self;
+    fn eq(self, other: Self) -> u8;
+    fn neq(self, other: Self) -> u8;
+    fn lt_eq(self, other: Self) -> u8;
+    fn lt(self, other: Self) -> u8;
+    fn gt(self, other: Self) -> u8;
+    fn gt_eq(self, other: Self) -> u8;
+}
+
+#[inline]
+pub(super) fn set<T: Copy, F: Fn(T, T) -> bool>(lhs: [T; 8], rhs: [T; 8], op: F) -> u8 {
+    let mut byte = 0u8;
+    lhs.iter()
+        .zip(rhs.iter())
+        .enumerate()
+        .for_each(|(i, (lhs, rhs))| {
+            byte |= if op(*lhs, *rhs) { 1 << i } else { 0 };
+        });
+    byte
+}
+
+macro_rules! simd8_native {
+    ($type:ty) => {
+        impl Simd8 for $type {
+            type Simd = [$type; 8];
+        }
+
+        impl Simd8Lanes<$type> for [$type; 8] {
+            #[inline]
+            fn from_chunk(v: &[$type]) -> Self {
+                v.try_into().unwrap()
+            }
+
+            #[inline]
+            fn from_incomplete_chunk(v: &[$type], remaining: $type) -> Self {
+                let mut a = [remaining; 8];
+                a.iter_mut().zip(v.iter()).for_each(|(a, b)| *a = *b);
+                a
+            }
+
+            #[inline]
+            fn eq(self, other: Self) -> u8 {
+                set(self, other, |x, y| x == y)
+            }
+
+            #[inline]
+            fn neq(self, other: Self) -> u8 {
+                #[allow(clippy::float_cmp)]
+                set(self, other, |x, y| x != y)
+            }
+
+            #[inline]
+            fn lt_eq(self, other: Self) -> u8 {
+                set(self, other, |x, y| x <= y)
+            }
+
+            #[inline]
+            fn lt(self, other: Self) -> u8 {
+                set(self, other, |x, y| x < y)
+            }
+
+            #[inline]
+            fn gt_eq(self, other: Self) -> u8 {
+                set(self, other, |x, y| x >= y)
+            }
+
+            #[inline]
+            fn gt(self, other: Self) -> u8 {
+                set(self, other, |x, y| x > y)
+            }
+        }
+    };
+}
+
+#[cfg(not(feature = "simd"))]
+mod native;
+#[cfg(not(feature = "simd"))]
+pub use native::*;
+#[cfg(feature = "simd")]
+mod packed;
+#[cfg(feature = "simd")]
+pub use packed::*;

src/compute/comparison/simd/native.rs

@@ -0,0 +1,15 @@
+use std::convert::TryInto;
+
+use super::{set, Simd8, Simd8Lanes};
+
+simd8_native!(u8);
+simd8_native!(u16);
+simd8_native!(u32);
+simd8_native!(u64);
+simd8_native!(i8);
+simd8_native!(i16);
+simd8_native!(i32);
+simd8_native!(i128);
+simd8_native!(i64);
+simd8_native!(f32);
+simd8_native!(f64);