test: use full range for floats uniform (jvdd#60)

* test: use full range for floats uniform

* 🧹 formatting

* tests: use SampleUniformFullRange in integration tests

* bench: update benchmarks with SampleUniformFullRange

* test: update int and uint with SampleUniformFullRange

* fix: use correct trait

* ci: skip MIPS as it is not supported on nightly anymore

* 🧹 remove unused function

* fix: use trait instead of removed function

* :detetctive: float uniform test

* ⬆️ update dev dependencies

* 🧹 fix arrow2 tests for half

* ⬆️ update codspeed CI action

Author: jvdd

.github/workflows/ci.yml

@@ -129,8 +129,12 @@ jobs:
           - i586-unknown-linux-gnu
           - aarch64-unknown-linux-gnu
           - armv7-unknown-linux-gnueabihf
-          - mips-unknown-linux-gnu
-          - mips64-unknown-linux-gnuabi64
+          # MIPS is currently not supported anymore on nightly chains.
+          # more information:
+          #   - https://github.com/rust-lang/compiler-team/issues/648
+          #   - https://github.com/rust-lang/rust/pull/113274
+          # - mips-unknown-linux-gnu
+          # - mips64-unknown-linux-gnuabi64
           - powerpc-unknown-linux-gnu
           - powerpc64-unknown-linux-gnu
           - riscv64gc-unknown-linux-gnu

.github/workflows/codspeed.yaml

@@ -43,7 +43,7 @@ jobs:
       - name: Build the benchmark target(s)
         run: cargo codspeed build --features half
 
-      - uses: CodSpeedHQ/action@v1
+      - uses: CodSpeedHQ/action@v2
         name: Run benchmarks
         with:
           run: cargo codspeed run

Cargo.toml

@@ -14,8 +14,8 @@ exclude = [".git*", "dev_utils/**/*", "tests/**/*"]
 
 
 [dependencies]
-num-traits = { version = "0.2.15", default-features = false }
-half = { version = "2.1.0", default-features = false, features=["num-traits"], optional = true }
+num-traits = { version = "0.2.17", default-features = false }
+half = { version = "2.3.1", default-features = false, features=["num-traits"], optional = true }
 ndarray = { version = "0.15.6", default-features = false, optional = true}
 arrow = { version = ">0", default-features = false, optional = true}
 arrow2 = { version = ">0.0", default-features = false, optional = true}
@@ -31,13 +31,16 @@ arrow = ["dep:arrow"]
 arrow2 = ["dep:arrow2"]
 
 [dev-dependencies]
-rstest = { version = "0.16", default-features = false }
-rstest_reuse = { version = "0.5", default-features = false }
-rand = { version = "0.7.2", default-features = false }
-codspeed-criterion-compat = "1.1"
-criterion = "0.3.1"
+rstest = { version = "0.18.2", default-features = false }
+rstest_reuse = { version = "0.6", default-features = false }
+rand = { version = "0.8.5", default-features = false }
+codspeed-criterion-compat = "2.3.3"
+criterion = "0.5.1"
 dev_utils = { path = "dev_utils" }
 
+[dev-dependencies.half]
+features = ["rand_distr"]
+
 
 [[bench]]
 name = "bench_f16_return_nan"

benches/bench_f16_ignore_nan.rs

@@ -13,28 +13,11 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 use half::f16;
 
-fn get_random_f16_array(n: usize) -> Vec<f16> {
-    let data = utils::get_random_array::<i16>(n, -0x7C00, 0x7C00);
-    let data: Vec<f16> = data.iter().map(|&x| f16::from_bits(x as u16)).collect();
-    // Replace NaNs and Infs with 0
-    let data: Vec<f16> = data
-        .iter()
-        .map(|&x| {
-            if x.is_nan() || x.is_infinite() {
-                f16::from_bits(0)
-            } else {
-                x
-            }
-        })
-        .collect();
-    data
-}
-
 // _in stands for "ignore nan"
 
 fn argminmax_in_f16_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f16] = &get_random_f16_array(n);
+    let data: &[f16] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f16_argminmax_in", |b| {
         b.iter(|| SCALAR::<FloatIgnoreNaN>::argminmax(black_box(data)))
     });

benches/bench_f16_return_nan.rs

@@ -13,28 +13,11 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 use half::f16;
 
-fn get_random_f16_array(n: usize) -> Vec<f16> {
-    let data = utils::get_random_array::<i16>(n, -0x7C00, 0x7C00);
-    let data: Vec<f16> = data.iter().map(|&x| f16::from_bits(x as u16)).collect();
-    // Replace NaNs and Infs with 0
-    let data: Vec<f16> = data
-        .iter()
-        .map(|&x| {
-            if x.is_nan() || x.is_infinite() {
-                f16::from_bits(0)
-            } else {
-                x
-            }
-        })
-        .collect();
-    data
-}
-
 // _rn stands for "return nan"
 
 fn argminmax_rn_f16_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f16] = &get_random_f16_array(n);
+    let data: &[f16] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f16_argminmax_rn", |b| {
         b.iter(|| SCALAR::<FloatReturnNaN>::argminmax(black_box(data)))
     });

benches/bench_f32_ignore_nan.rs

@@ -15,7 +15,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_in_f32_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f32] = &utils::get_random_array::<f32>(n, f32::MIN, f32::MAX);
+    let data: &[f32] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f32_argminmax_in", |b| {
         b.iter(|| SCALAR::<FloatIgnoreNaN>::argminmax(black_box(data)))
     });

benches/bench_f32_return_nan.rs

@@ -15,7 +15,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_rn_f32_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f32] = &utils::get_random_array::<f32>(n, f32::MIN, f32::MAX);
+    let data: &[f32] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f32_argminmax_rn", |b| {
         b.iter(|| SCALAR::<FloatReturnNaN>::argminmax(black_box(data)))
     });

benches/bench_f64_ignore_nan.rs

@@ -15,7 +15,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_in_f64_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f64] = &utils::get_random_array::<f64>(n, f64::MIN, f64::MAX);
+    let data: &[f64] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f64_argminmax_in", |b| {
         b.iter(|| SCALAR::<FloatIgnoreNaN>::argminmax(black_box(data)))
     });

benches/bench_f64_return_nan.rs

@@ -15,7 +15,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_rn_f64_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[f64] = &utils::get_random_array::<f64>(n, f64::MIN, f64::MAX);
+    let data: &[f64] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_f64_argminmax_rn", |b| {
         b.iter(|| SCALAR::<FloatReturnNaN>::argminmax(black_box(data)))
     });

benches/bench_i16.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_i16_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[i16] = &utils::get_random_array::<i16>(n, i16::MIN, i16::MAX);
+    let data: &[i16] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_i16_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_i32.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_i32_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[i32] = &utils::get_random_array::<i32>(n, i32::MIN, i32::MAX);
+    let data: &[i32] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_i32_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_i64.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_i64_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[i64] = &utils::get_random_array::<i64>(n, i64::MIN, i64::MAX);
+    let data: &[i64] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_i64_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_i8.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_i8_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[i8] = &utils::get_random_array::<i8>(n, i8::MIN, i8::MAX);
+    let data: &[i8] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_i8_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_u16.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_u16_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[u16] = &utils::get_random_array::<u16>(n, u16::MIN, u16::MAX);
+    let data: &[u16] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_u16_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_u32.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_u32_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[u32] = &utils::get_random_array::<u32>(n, u32::MIN, u32::MAX);
+    let data: &[u32] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_u32_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_u64.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_u64_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[u64] = &utils::get_random_array::<u64>(n, u64::MIN, u64::MAX);
+    let data: &[u64] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_u64_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

benches/bench_u8.rs

@@ -12,7 +12,7 @@ use argminmax::simd::{SIMDArgMinMax, NEON};
 
 fn argminmax_u8_random_array_long(c: &mut Criterion) {
     let n = config::ARRAY_LENGTH_LONG;
-    let data: &[u8] = &utils::get_random_array::<u8>(n, u8::MIN, u8::MAX);
+    let data: &[u8] = &utils::SampleUniformFullRange::get_random_array(n);
     c.bench_function("scalar_u8_argminmax", |b| {
         b.iter(|| SCALAR::argminmax(black_box(data)))
     });

dev_utils/Cargo.toml

@@ -6,5 +6,6 @@ edition = "2021"
 description = "Shared utilities for development (tests & benchmarks)"
 
 [dependencies]
-rand = { version = "0.7.2", default-features = false }
-rand_distr = { version = "0.2.2", default-features = false }
+num-traits = { version = "0.2.17", default-features = false }
+half = { version = "2.3.1", default-features = false, features = ["num-traits", "rand_distr"] }
+rand = { version = "0.8.5" }

dev_utils/src/utils.rs

@@ -1,17 +1,10 @@
+use half::f16;
+use num_traits::Zero;
+
 use std::ops::{Add, Sub};
 
+use rand::distributions::Uniform;
 use rand::{thread_rng, Rng};
-use rand_distr::Uniform;
-
-// random array that samples between min and max of T
-pub fn get_random_array<T>(n: usize, min_value: T, max_value: T) -> Vec<T>
-where
-    T: Copy + rand::distributions::uniform::SampleUniform,
-{
-    let rng = thread_rng();
-    let uni = Uniform::new_inclusive(min_value, max_value);
-    rng.sample_iter(uni).take(n).collect()
-}
 
 // worst case array that alternates between increasing max and decreasing min values
 pub fn get_worst_case_array<T>(n: usize, step: T) -> Vec<T>
@@ -32,3 +25,60 @@ where
     }
     arr
 }
+
+pub trait SampleUniformFullRange: rand::distributions::uniform::SampleUniform {
+    const MIN: Self;
+    const MAX: Self;
+
+    // random array that samples between min and max of Self
+    fn get_random_array(n: usize) -> Vec<Self>
+    where
+        Self: Copy + rand::distributions::uniform::SampleUniform,
+    {
+        let rng = thread_rng();
+        let uni = Uniform::new_inclusive(Self::MIN, Self::MAX);
+        rng.sample_iter(uni).take(n).collect()
+    }
+}
+
+macro_rules! impl_full_range_uniform {
+    ($($t:ty),*) => {
+        $(
+            impl SampleUniformFullRange for $t {
+                const MIN: Self = <$t>::MIN;
+                const MAX: Self = <$t>::MAX;
+            }
+        )*
+    };
+}
+
+macro_rules! impl_full_range_uniform_float {
+    ($($t:ty, $t_int:ty),*) => {
+        $(
+            impl SampleUniformFullRange for $t {
+                // These 2 are not used, but are required by the trait
+                const MIN: Self = <$t>::MIN;
+                const MAX: Self = <$t>::MAX;
+
+                fn get_random_array(n: usize) -> Vec<Self>
+                where
+                    Self: Copy + rand::distributions::uniform::SampleUniform,
+                {
+                    // Get a uniform random array of integers
+                    let rand_arr_int: Vec<$t_int> = <$t_int>::get_random_array(n);
+                    // Transmute the integers to floats
+                    let rand_arr_float: Vec<Self> = unsafe { std::mem::transmute(rand_arr_int) };
+                    // Replace the NaNs with 0.0
+                    rand_arr_float.iter().map(|x| if x.is_nan() { <$t>::zero() } else { *x }).collect()
+                }
+            }
+        )*
+    };
+}
+
+impl_full_range_uniform!(i8, i16, i32, i64, u8, u16, u32, u64);
+// f16 does not suffer from range overflow panick as upcast to f32 is used to generate
+// the random numbers
+impl_full_range_uniform!(f16);
+// Workaround for f32 and f64 as these suffer from range overflow in the rand crate
+impl_full_range_uniform_float!(f32, i32, f64, i64);

src/scalar/scalar_f16.rs

@@ -238,10 +238,8 @@ mod tests {
     const ARR_LEN: usize = 1025;
 
     fn get_arrays(len: usize) -> (Vec<f32>, Vec<f16>) {
-        // we use i8 its to make sure we have correct representation in float
-        let v = utils::get_random_array(len, i8::MIN, i8::MAX);
-        let vec_f32: Vec<f32> = v.iter().map(|x| *x as f32).collect();
-        let vec_f16: Vec<f16> = vec_f32.iter().map(|x| f16::from_f32(*x)).collect();
+        let vec_f16: Vec<f16> = utils::SampleUniformFullRange::get_random_array(len);
+        let vec_f32: Vec<f32> = vec_f16.iter().map(|x| x.to_f32()).collect();
         (vec_f32, vec_f16)
     }
 

src/simd/simd_f16_ignore_nan.rs

@@ -726,8 +726,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f16(n: usize) -> Vec<f16> {
-        let arr = utils::get_random_array(n, i16::MIN, i16::MAX);
-        arr.iter().map(|x| f16::from_f32(*x as f32)).collect()
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_f16_return_nan.rs

@@ -687,8 +687,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f16(n: usize) -> Vec<f16> {
-        let arr = utils::get_random_array(n, i16::MIN, i16::MAX);
-        arr.iter().map(|x| f16::from_f32(*x as f32)).collect()
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_f32_ignore_nan.rs

@@ -355,7 +355,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f32(n: usize) -> Vec<f32> {
-        utils::get_random_array(n, f32::MIN, f32::MAX)
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_f32_return_nan.rs

@@ -501,7 +501,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f32(n: usize) -> Vec<f32> {
-        utils::get_random_array(n, f32::MIN, f32::MAX)
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_f64_ignore_nan.rs

@@ -377,7 +377,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f64(n: usize) -> Vec<f64> {
-        utils::get_random_array(n, f64::MIN, f64::MAX)
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_f64_return_nan.rs

@@ -552,7 +552,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_f64(n: usize) -> Vec<f64> {
-        utils::get_random_array(n, f64::MIN, f64::MAX)
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation

src/simd/simd_i16.rs

@@ -538,7 +538,7 @@ mod tests {
     use dev_utils::utils;
 
     fn get_array_i16(n: usize) -> Vec<i16> {
-        utils::get_random_array(n, i16::MIN, i16::MAX)
+        utils::SampleUniformFullRange::get_random_array(n)
     }
 
     // The scalar implementation