Feature/round

.github/workflows/ci.yml

@@ -121,7 +121,8 @@ jobs:
           # for NaNs which makes it worth testing on despite that.
           - mips-unknown-linux-gnu
           - mips64-unknown-linux-gnuabi64
-          - riscv64gc-unknown-linux-gnu
+          # TODO: reenable pending https://github.com/rust-lang/rust/issues/77866
+          # - riscv64gc-unknown-linux-gnu
 
     steps:
       - uses: actions/checkout@v2

crates/core_simd/src/intrinsics.rs

@@ -36,4 +36,7 @@ extern "platform-intrinsic" {
 
     /// xor
     pub(crate) fn simd_xor<T>(x: T, y: T) -> T;
+
+    /// fptoui/fptosi/uitofp/sitofp
+    pub(crate) fn simd_cast<T, U>(x: T) -> U;
 }

crates/core_simd/src/lib.rs

@@ -1,5 +1,5 @@
 #![no_std]
-#![feature(repr_simd, platform_intrinsics)]
+#![feature(repr_simd, platform_intrinsics, link_llvm_intrinsics, simd_ffi)]
 #![warn(missing_docs)]
 //! Portable SIMD module.
 
@@ -56,3 +56,5 @@ mod vectors_mask128;
 pub use vectors_mask128::*;
 mod vectors_masksize;
 pub use vectors_masksize::*;
+
+mod round;

crates/core_simd/src/round.rs

@@ -0,0 +1,153 @@
+macro_rules! implement {
+    {
+        impl $type:ident {
+            int_type = $int_type:ident,
+            floor = $floor_intrinsic:literal,
+            ceil = $ceil_intrinsic:literal,
+            round = $round_intrinsic:literal,
+            trunc = $trunc_intrinsic:literal,
+        }
+    } => {
+        mod $type {
+            #[allow(improper_ctypes)]
+            extern "C" {
+                #[link_name = $floor_intrinsic]
+                fn floor_intrinsic(x: crate::$type) -> crate::$type;
+                #[link_name = $ceil_intrinsic]
+                fn ceil_intrinsic(x: crate::$type) -> crate::$type;
+                #[link_name = $round_intrinsic]
+                fn round_intrinsic(x: crate::$type) -> crate::$type;
+                #[link_name = $trunc_intrinsic]
+                fn trunc_intrinsic(x: crate::$type) -> crate::$type;
+            }
+
+            impl crate::$type {
+                /// Returns the largest integer less than or equal to each lane.
+                #[must_use = "method returns a new vector and does not mutate the original value"]
+                #[inline]
+                pub fn floor(self) -> Self {
+                    unsafe { floor_intrinsic(self) }
+                }
+
+                /// Returns the smallest integer greater than or equal to each lane.
+                #[must_use = "method returns a new vector and does not mutate the original value"]
+                #[inline]
+                pub fn ceil(self) -> Self {
+                    unsafe { ceil_intrinsic(self) }
+                }
+
+                /// Returns the nearest integer to each lane. Round half-way cases away from 0.0.
+                #[must_use = "method returns a new vector and does not mutate the original value"]
+                #[inline]
+                pub fn round(self) -> Self {
+                    unsafe { round_intrinsic(self) }
+                }
+
+                /// Returns the integer part of each lane.
+                #[must_use = "method returns a new vector and does not mutate the original value"]
+                #[inline]
+                pub fn trunc(self) -> Self {
+                    unsafe { trunc_intrinsic(self) }
+                }
+
+                /// Returns the fractional part of each lane.
+                #[must_use = "method returns a new vector and does not mutate the original value"]
+                #[inline]
+                pub fn fract(self) -> Self {
+                    self - self.trunc()
+                }
+
+                /// Rounds toward zero and converts to the same-width integer type, assuming that
+                /// the value is finite and fits in that type.
+                ///
+                /// # Safety
+                /// The value must:
+                ///
+                /// * Not be NaN
+                /// * Not be infinite
+                /// * Be representable in the return type, after truncating off its fractional part
+                #[inline]
+                pub unsafe fn to_int_unchecked(self) -> crate::$int_type {
+                    crate::intrinsics::simd_cast(self)
+                }
+
+                /// Creates a floating-point vector from an integer vector.  Rounds values that are
+                /// not exactly representable.
+                #[inline]
+                pub fn round_from_int(value: crate::$int_type) -> Self {
+                    unsafe { crate::intrinsics::simd_cast(value) }
+                }
+            }
+        }
+    }
+}
+
+implement! {
+    impl f32x2 {
+        int_type = i32x2,
+        floor = "llvm.floor.v2f32",
+        ceil = "llvm.ceil.v2f32",
+        round = "llvm.round.v2f32",
+        trunc = "llvm.trunc.v2f32",
+    }
+}
+
+implement! {
+    impl f32x4 {
+        int_type = i32x4,
+        floor = "llvm.floor.v4f32",
+        ceil = "llvm.ceil.v4f32",
+        round = "llvm.round.v4f32",
+        trunc = "llvm.trunc.v4f32",
+    }
+}
+
+implement! {
+    impl f32x8 {
+        int_type = i32x8,
+        floor = "llvm.floor.v8f32",
+        ceil = "llvm.ceil.v8f32",
+        round = "llvm.round.v8f32",
+        trunc = "llvm.trunc.v8f32",
+    }
+}
+
+implement! {
+    impl f32x16 {
+        int_type = i32x16,
+        floor = "llvm.floor.v16f32",
+        ceil = "llvm.ceil.v16f32",
+        round = "llvm.round.v16f32",
+        trunc = "llvm.trunc.v16f32",
+    }
+}
+
+implement! {
+    impl f64x2 {
+        int_type = i64x2,
+        floor = "llvm.floor.v2f64",
+        ceil = "llvm.ceil.v2f64",
+        round = "llvm.round.v2f64",
+        trunc = "llvm.trunc.v2f64",
+    }
+}
+
+implement! {
+    impl f64x4 {
+        int_type = i64x4,
+        floor = "llvm.floor.v4f64",
+        ceil = "llvm.ceil.v4f64",
+        round = "llvm.round.v4f64",
+        trunc = "llvm.trunc.v4f64",
+    }
+}
+
+implement! {
+    impl f64x8 {
+        int_type = i64x8,
+        floor = "llvm.floor.v8f64",
+        ceil = "llvm.ceil.v8f64",
+        round = "llvm.round.v8f64",
+        trunc = "llvm.trunc.v8f64",
+    }
+}

crates/core_simd/tests/helpers/lanewise.rs

@@ -1,8 +1,13 @@
-pub fn apply_unary_lanewise<T: Copy, V: AsMut<[T]> + Default>(mut x: V, f: impl Fn(T) -> T) -> V {
-    for lane in x.as_mut() {
-        *lane = f(*lane)
+pub fn apply_unary_lanewise<T1: Copy, T2: Copy, V1: AsRef<[T1]>, V2: AsMut<[T2]> + Default>(
+    x: V1,
+    f: impl Fn(T1) -> T2,
+) -> V2 {
+    let mut y = V2::default();
+    assert_eq!(x.as_ref().len(), y.as_mut().len());
+    for (x, y) in x.as_ref().iter().zip(y.as_mut().iter_mut()) {
+        *y = f(*x);
     }
-    x
+    y
 }
 
 pub fn apply_binary_lanewise<T: Copy, V: AsRef<[T]> + AsMut<[T]> + Default>(

crates/core_simd/tests/ops_impl/f32.rs

@@ -1,6 +1,6 @@
 use super::helpers;
 
-float_tests! { f32x2, f32 }
-float_tests! { f32x4, f32 }
-float_tests! { f32x8, f32 }
-float_tests! { f32x16, f32 }
+float_tests! { f32x2, f32, i32x2, i32 }
+float_tests! { f32x4, f32, i32x4, i32 }
+float_tests! { f32x8, f32, i32x8, i32 }
+float_tests! { f32x16, f32, i32x16, i32 }

crates/core_simd/tests/ops_impl/f64.rs

@@ -1,5 +1,5 @@
 use super::helpers;
 
-float_tests! { f64x2, f64 }
-float_tests! { f64x4, f64 }
-float_tests! { f64x8, f64 }
+float_tests! { f64x2, f64, i64x2, i64 }
+float_tests! { f64x4, f64, i64x4, i64 }
+float_tests! { f64x8, f64, i64x8, i64 }

crates/core_simd/tests/ops_impl/float_macros.rs

@@ -1,5 +1,5 @@
 macro_rules! float_tests {
-    { $vector:ident, $scalar:ident } => {
+    { $vector:ident, $scalar:ident, $int_vector:ident, $int_scalar:ident } => {
         #[cfg(test)]
         mod $vector {
             use super::*;
@@ -24,6 +24,18 @@ macro_rules! float_tests {
                 slice.chunks_exact(lanes).map(from_slice)
             }
 
+            fn from_slice_int(slice: &[$int_scalar]) -> core_simd::$int_vector {
+                let mut value = core_simd::$int_vector::default();
+                let value_slice: &mut [_] = value.as_mut();
+                value_slice.copy_from_slice(&slice[0..value_slice.len()]);
+                value
+            }
+
+            fn slice_chunks_int(slice: &[$int_scalar]) -> impl Iterator<Item = core_simd::$int_vector> + '_ {
+                let lanes = core::mem::size_of::<core_simd::$int_vector>() / core::mem::size_of::<$int_scalar>();
+                slice.chunks_exact(lanes).map(from_slice_int)
+            }
+
             const A: [$scalar; 16] = [0.,   1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9., 10., 11., 12., 13., 14., 15.];
             const B: [$scalar; 16] = [16., 17., 18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31.];
             const C: [$scalar; 16] = [
@@ -322,6 +334,121 @@ macro_rules! float_tests {
                     assert_biteq!(v.abs(), expected);
                 }
             }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn ceil_odd_floats() {
+                for v in slice_chunks(&C) {
+                    let expected = apply_unary_lanewise(v, <$scalar>::ceil);
+                    assert_biteq!(v.ceil(), expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn floor_odd_floats() {
+                for v in slice_chunks(&C) {
+                    let expected = apply_unary_lanewise(v, <$scalar>::floor);
+                    assert_biteq!(v.floor(), expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn round_odd_floats() {
+                for v in slice_chunks(&C) {
+                    let expected = apply_unary_lanewise(v, <$scalar>::round);
+                    assert_biteq!(v.round(), expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn round_mode() {
+                assert_biteq!(core_simd::$vector::splat(1.5).round(), core_simd::$vector::splat(2.0));
+                assert_biteq!(core_simd::$vector::splat(2.5).round(), core_simd::$vector::splat(3.0));
+                assert_biteq!(core_simd::$vector::splat(-1.5).round(), core_simd::$vector::splat(-2.0));
+                assert_biteq!(core_simd::$vector::splat(-2.5).round(), core_simd::$vector::splat(-3.0));
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn trunc_odd_floats() {
+                for v in slice_chunks(&C) {
+                    let expected = apply_unary_lanewise(v, <$scalar>::trunc);
+                    assert_biteq!(v.trunc(), expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn fract_odd_floats() {
+                for v in slice_chunks(&C) {
+                    let expected = apply_unary_lanewise(v, <$scalar>::fract);
+                    assert_biteq!(v.fract(), expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn to_int_unchecked() {
+                // The maximum integer that can be represented by the equivalently sized float has
+                // all of the mantissa digits set to 1, pushed up to the MSB.
+                const ALL_MANTISSA_BITS: $int_scalar = ((1 << <$scalar>::MANTISSA_DIGITS) - 1);
+                const MAX_REPRESENTABLE_VALUE: $int_scalar =
+                    ALL_MANTISSA_BITS << (core::mem::size_of::<$scalar>() * 8 - <$scalar>::MANTISSA_DIGITS as usize - 1);
+                const VALUES: [$scalar; 16] = [
+                    -0.0,
+                    0.0,
+                    -1.0,
+                    1.0,
+                    ALL_MANTISSA_BITS as $scalar,
+                    -ALL_MANTISSA_BITS as $scalar,
+                    MAX_REPRESENTABLE_VALUE as $scalar,
+                    -MAX_REPRESENTABLE_VALUE as $scalar,
+                    (MAX_REPRESENTABLE_VALUE / 2) as $scalar,
+                    (-MAX_REPRESENTABLE_VALUE / 2) as $scalar,
+                    <$scalar>::MIN_POSITIVE,
+                    -<$scalar>::MIN_POSITIVE,
+                    <$scalar>::EPSILON,
+                    -<$scalar>::EPSILON,
+                    100.0 / 3.0,
+                    -100.0 / 3.0,
+                ];
+
+                for v in slice_chunks(&VALUES) {
+                    let expected = apply_unary_lanewise(v, |x| unsafe { x.to_int_unchecked() });
+                    assert_biteq!(unsafe { v.to_int_unchecked() }, expected);
+                }
+            }
+
+            #[test]
+            #[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
+            fn round_from_int() {
+                const VALUES: [$int_scalar; 16] = [
+                    0,
+                    0,
+                    1,
+                    -1,
+                    100,
+                    -100,
+                    200,
+                    -200,
+                    413,
+                    -413,
+                    1017,
+                    -1017,
+                    1234567,
+                    -1234567,
+                    <$int_scalar>::MAX,
+                    <$int_scalar>::MIN,
+                ];
+
+                for v in slice_chunks_int(&VALUES) {
+                    let expected = apply_unary_lanewise(v, |x| x as $scalar);
+                    assert_biteq!(core_simd::$vector::round_from_int(v), expected);
+                }
+            }
         }
     }
 }