vendor arrow simd operations so that we can use latest nightly

README.md

@@ -197,7 +197,7 @@ Additional cargo features:
     - pretty printing of DataFrames
 * `temporal (default)`
     - Conversions between Chrono and Polars for temporal data
-* `simd`
+* `simd (default)`
     - SIMD operations
 * `parquet`
     - Read Apache Parquet format

polars/Cargo.toml

@@ -12,12 +12,12 @@ readme = "../README.md"
 
 [features]
 pretty = ["prettytable-rs"]
-simd = ["arrow/simd"]
+simd = [] #["arrow/simd"]
 docs = []
 temporal = ["chrono"]
 random = ["rand", "rand_distr"]
 parallel = []
-default = ["pretty", "docs", "temporal"]
+default = ["pretty", "docs", "temporal", "simd"]
 lazy = []
 
 [dependencies]
@@ -33,4 +33,5 @@ chrono = {version = "^0.4.13", optional = true}
 parquet = {version = "1", optional = true}
 rand = {version = "0.7", optional = true}
 rand_distr = {version = "0.3", optional = true}
-ndarray = {version = "0.13", optional = true, default_features = false}
+ndarray = {version = "0.13", optional = true, default_features = false}
+packed_simd_2 = "0.3.4"

polars/src/chunked_array/kernels/vendor/arithmetic.rs

@@ -34,13 +34,13 @@ use num::{One, Zero};
 use super::utils::apply_bin_op_to_option_bitmap;
 #[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
 use super::utils::simd_load_set_invalid;
+use crate::datatypes::PolarsNumericType;
 use arrow::array::*;
 #[cfg(feature = "simd")]
 use arrow::bitmap::Bitmap;
 use arrow::buffer::Buffer;
 #[cfg(feature = "simd")]
 use arrow::buffer::MutableBuffer;
-use arrow::datatypes;
 use arrow::datatypes::ToByteSlice;
 use arrow::error::{ArrowError, Result};
 use arrow::util::bit_util;
@@ -54,7 +54,7 @@ pub fn math_op<T, F>(
     op: F,
 ) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     F: Fn(T::Native, T::Native) -> Result<T::Native>,
 {
     if left.len() != right.len() {
@@ -106,7 +106,7 @@ fn simd_math_op<T, F>(
     op: F,
 ) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Simd: Add<Output = T::Simd>
         + Sub<Output = T::Simd>
         + Mul<Output = T::Simd>
@@ -161,7 +161,7 @@ where
 #[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
 fn simd_divide<T>(left: &PrimitiveArray<T>, right: &PrimitiveArray<T>) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: One + Zero,
 {
     if left.len() != right.len() {
@@ -203,7 +203,7 @@ where
         T::write(simd_result, result_slice);
     }
 
-    let null_bit_buffer = bitmap.map(|b| b.bits);
+    let null_bit_buffer = bitmap.map(|b| b.into_buffer());
 
     let data = ArrayData::new(
         T::get_data_type(),
@@ -221,7 +221,7 @@ where
 /// then the result is also null.
 pub fn add<T>(left: &PrimitiveArray<T>, right: &PrimitiveArray<T>) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: Add<Output = T::Native>
         + Sub<Output = T::Native>
         + Mul<Output = T::Native>
@@ -239,7 +239,7 @@ where
 /// then the result is also null.
 pub fn subtract<T>(left: &PrimitiveArray<T>, right: &PrimitiveArray<T>) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: Add<Output = T::Native>
         + Sub<Output = T::Native>
         + Mul<Output = T::Native>
@@ -257,7 +257,7 @@ where
 /// then the result is also null.
 pub fn multiply<T>(left: &PrimitiveArray<T>, right: &PrimitiveArray<T>) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: Add<Output = T::Native>
         + Sub<Output = T::Native>
         + Mul<Output = T::Native>
@@ -276,7 +276,7 @@ where
 /// operation will be `Err(ArrowError::DivideByZero)`.
 pub fn divide<T>(left: &PrimitiveArray<T>, right: &PrimitiveArray<T>) -> Result<PrimitiveArray<T>>
 where
-    T: datatypes::ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: Add<Output = T::Native>
         + Sub<Output = T::Native>
         + Mul<Output = T::Native>

polars/src/chunked_array/kernels/vendor/utils.rs

@@ -1,13 +1,14 @@
+#[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
+use crate::datatypes::PolarsNumericType;
+#[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
+use arrow::array::PrimitiveArray;
 use arrow::bitmap::Bitmap;
 use arrow::buffer::Buffer;
 use arrow::error::Result;
 #[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
-use arrow::{
-    array::{ArrayDataRef, PrimitiveArray},
-    datatypes::ArrowNumericType,
-};
-#[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
 use num::One;
+#[cfg(feature = "simd")]
+use std::cmp::min;
 
 /// Performs a SIMD load but sets all 'invalid' lanes to a constant value.
 ///
@@ -26,7 +27,7 @@ pub(crate) unsafe fn simd_load_set_invalid<T>(
     fill_value: T::Native,
 ) -> T::Simd
 where
-    T: ArrowNumericType,
+    T: PolarsNumericType,
     T::Native: One,
 {
     let simd_with_zeros = T::load(array.value_slice(i, simd_width));
@@ -59,3 +60,39 @@ where
         },
     }
 }
+
+/// Creates a new SIMD mask, i.e. `packed_simd::m32x16` or similar. that indicates if the
+/// corresponding array slots represented by the mask are 'valid'.
+///
+/// Lanes of the SIMD mask can be set to 'valid' (`true`) if the corresponding array slot is not
+/// `NULL`, as indicated by it's `Bitmap`, and is within the length of the array.  Lanes outside the
+/// length represent padding and are set to 'invalid' (`false`).
+#[cfg(all(any(target_arch = "x86", target_arch = "x86_64"), feature = "simd"))]
+unsafe fn is_valid<T>(
+    bitmap: &Option<Bitmap>,
+    i: usize,
+    simd_width: usize,
+    array_len: usize,
+) -> T::SimdMask
+where
+    T: PolarsNumericType,
+{
+    let simd_upper_bound = i + simd_width;
+    let mut validity = T::mask_init(true);
+
+    // Validity based on `Bitmap`
+    if let Some(b) = bitmap {
+        for j in i..min(array_len, simd_upper_bound) {
+            if !b.is_set(j) {
+                validity = T::mask_set(validity, j - i, false);
+            }
+        }
+    }
+
+    // Validity based on the length of the Array
+    for j in array_len..simd_upper_bound {
+        validity = T::mask_set(validity, j - i, false);
+    }
+
+    validity
+}