Add unsigned_abs to i8,i16,i32,i64 to provide well defined behavior f…

…or abs (#152) * add unsigned version of abs for clearer code when handling corner cases * fix typo * fixed wasm typo * fix neon cast * add pub(crate) to unsigned types * fixed i8x16 * x86 doesnt have abs i64 * typo * fix 64 bit on intel * fix abs * remove trait accidentally added during merge * replace map with explicit array after reading warning about bad debug perf
Lokathor · May 12, 2024 · 445451e · 445451e
1 parent 4875c14
commit 445451e
Show file tree

Hide file tree

Showing 18 changed files with 318 additions and 18 deletions.
diff --git a/src/i16x8_.rs b/src/i16x8_.rs
@@ -696,10 +696,52 @@ impl i16x8 {
       } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
         unsafe {Self { neon: vabsq_s16(self.neon) }}
       } else {
-        self.is_negative().blend(self.neg(), self)
+        let arr: [i16; 8] = cast(self);
+        cast(
+          [
+            arr[0].wrapping_abs(),
+            arr[1].wrapping_abs(),
+            arr[2].wrapping_abs(),
+            arr[3].wrapping_abs(),
+            arr[4].wrapping_abs(),
+            arr[5].wrapping_abs(),
+            arr[6].wrapping_abs(),
+            arr[7].wrapping_abs(),
+          ])
+      }
+    }
+  }
+
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u16x8 {
+    pick! {
+      if #[cfg(target_feature="sse2")] {
+        let mask = shr_imm_i16_m128i::<15>(self.sse);
+        u16x8 { sse: bitxor_m128i(add_i16_m128i(self.sse, mask), mask) }
+      } else if #[cfg(target_feature="ssse3")] {
+        u16x8 { sse: abs_i16_m128i(self.sse) }
+      } else if #[cfg(target_feature="simd128")] {
+        u16x8 { simd: i16x8_abs(self.simd) }
+      } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
+        unsafe {u16x8 { neon: vreinterpretq_u16_s16(vabsq_s16(self.neon)) }}
+      } else {
+        let arr: [i16; 8] = cast(self);
+        cast(
+          [
+            arr[0].unsigned_abs(),
+            arr[1].unsigned_abs(),
+            arr[2].unsigned_abs(),
+            arr[3].unsigned_abs(),
+            arr[4].unsigned_abs(),
+            arr[5].unsigned_abs(),
+            arr[6].unsigned_abs(),
+            arr[7].unsigned_abs(),
+          ])
       }
     }
   }
+
   #[inline]
   #[must_use]
   pub fn max(self, rhs: Self) -> Self {

diff --git a/src/i32x4_.rs b/src/i32x4_.rs
@@ -439,6 +439,28 @@ impl i32x4 {
     }
   }
 
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u32x4 {
+    pick! {
+      if #[cfg(target_feature="ssse3")] {
+        u32x4 { sse: abs_i32_m128i(self.sse) }
+      } else if #[cfg(target_feature="simd128")] {
+        u32x4 { simd: i32x4_abs(self.simd) }
+      } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
+        unsafe {u32x4 { neon: vreinterpretq_u32_s32(vabsq_s32(self.neon)) }}
+      } else {
+        let arr: [i32; 4] = cast(self);
+        cast([
+          arr[0].unsigned_abs(),
+          arr[1].unsigned_abs(),
+          arr[2].unsigned_abs(),
+          arr[3].unsigned_abs(),
+        ])
+      }
+    }
+  }
+
   /// horizontal add of all the elements of the vector
   #[inline]
   #[must_use]

diff --git a/src/i32x8_.rs b/src/i32x8_.rs
@@ -369,6 +369,22 @@ impl i32x8 {
       }
     }
   }
+
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u32x8 {
+    pick! {
+      if #[cfg(target_feature="avx2")] {
+        u32x8 { avx2: abs_i32_m256i(self.avx2) }
+      } else {
+        u32x8 {
+          a : self.a.unsigned_abs(),
+          b : self.b.unsigned_abs(),
+        }
+      }
+    }
+  }
+
   #[inline]
   #[must_use]
   pub fn max(self, rhs: Self) -> Self {

diff --git a/src/i64x2_.rs b/src/i64x2_.rs
@@ -398,6 +398,46 @@ impl i64x2 {
     }
   }
 
+  #[inline]
+  #[must_use]
+  pub fn abs(self) -> Self {
+    pick! {
+      // x86 doesn't have this builtin
+      if #[cfg(target_feature="simd128")] {
+        Self { simd: i64x2_abs(self.simd) }
+      } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
+        unsafe {Self { neon: vabsq_s64(self.neon) }}
+      } else {
+        let arr: [i64; 2] = cast(self);
+        cast(
+          [
+            arr[0].wrapping_abs(),
+            arr[1].wrapping_abs(),
+          ])
+      }
+    }
+  }
+
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u64x2 {
+    pick! {
+      // x86 doesn't have this builtin
+      if #[cfg(target_feature="simd128")] {
+        u64x2 { simd: i64x2_abs(self.simd) }
+      } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
+        unsafe {u64x2 { neon: vreinterpretq_u64_s64(vabsq_s64(self.neon)) }}
+      } else {
+        let arr: [i64; 2] = cast(self);
+        cast(
+          [
+            arr[0].unsigned_abs(),
+            arr[1].unsigned_abs(),
+          ])
+      }
+    }
+  }
+
   #[inline]
   #[must_use]
   pub fn round_float(self) -> f64x2 {

diff --git a/src/i64x4_.rs b/src/i64x4_.rs
@@ -308,6 +308,52 @@ impl i64x4 {
     }
   }
 
+  #[inline]
+  #[must_use]
+  pub fn abs(self) -> Self {
+    pick! {
+      if #[cfg(target_feature="avx2")] {
+        // avx x86 doesn't have this builtin
+        let arr: [i64; 4] = cast(self);
+        cast(
+          [
+            arr[0].wrapping_abs(),
+            arr[1].wrapping_abs(),
+            arr[2].wrapping_abs(),
+            arr[3].wrapping_abs(),
+          ])
+      } else {
+        Self {
+          a : self.a.abs(),
+          b : self.b.abs(),
+        }
+      }
+    }
+  }
+
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u64x4 {
+    pick! {
+      if #[cfg(target_feature="avx2")] {
+        // avx x86 doesn't have this builtin
+        let arr: [i64; 4] = cast(self);
+        cast(
+          [
+            arr[0].unsigned_abs(),
+            arr[1].unsigned_abs(),
+            arr[2].unsigned_abs(),
+            arr[3].unsigned_abs(),
+          ])
+      } else {
+        u64x4 {
+          a : self.a.unsigned_abs(),
+          b : self.b.unsigned_abs(),
+        }
+      }
+    }
+  }
+
   #[inline]
   #[must_use]
   pub fn round_float(self) -> f64x4 {

diff --git a/src/i8x16_.rs b/src/i8x16_.rs
@@ -527,6 +527,42 @@ impl i8x16 {
       }
     }
   }
+
+  #[inline]
+  #[must_use]
+  pub fn unsigned_abs(self) -> u8x16 {
+    pick! {
+      if #[cfg(target_feature="ssse3")] {
+        u8x16 { sse: abs_i8_m128i(self.sse) }
+      } else if #[cfg(target_feature="simd128")] {
+        u8x16 { simd: i8x16_abs(self.simd) }
+      } else if #[cfg(all(target_feature="neon",target_arch="aarch64"))]{
+        unsafe { u8x16 { neon: vreinterpretq_u8_s8(vabsq_s8(self.neon)) }}
+      } else {
+        let arr: [i8; 16] = cast(self);
+        cast(
+          [
+            arr[0].unsigned_abs(),
+            arr[1].unsigned_abs(),
+            arr[2].unsigned_abs(),
+            arr[3].unsigned_abs(),
+            arr[4].unsigned_abs(),
+            arr[5].unsigned_abs(),
+            arr[6].unsigned_abs(),
+            arr[7].unsigned_abs(),
+            arr[8].unsigned_abs(),
+            arr[9].unsigned_abs(),
+            arr[10].unsigned_abs(),
+            arr[11].unsigned_abs(),
+            arr[12].unsigned_abs(),
+            arr[13].unsigned_abs(),
+            arr[14].unsigned_abs(),
+            arr[15].unsigned_abs(),
+            ])
+      }
+    }
+  }
+
   #[inline]
   #[must_use]
   pub fn max(self, rhs: Self) -> Self {

diff --git a/src/u16x8_.rs b/src/u16x8_.rs
@@ -4,13 +4,13 @@ pick! {
   if #[cfg(target_feature="sse2")] {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(16))]
-    pub struct u16x8 { sse: m128i }
+    pub struct u16x8 { pub(crate) sse: m128i }
   } else if #[cfg(target_feature="simd128")] {
     use core::arch::wasm32::*;
 
     #[derive(Clone, Copy)]
     #[repr(transparent)]
-    pub struct u16x8 { simd: v128 }
+    pub struct u16x8 { pub(crate) simd: v128 }
 
     impl Default for u16x8 {
       fn default() -> Self {
@@ -29,7 +29,7 @@ pick! {
       use core::arch::aarch64::*;
       #[repr(C)]
       #[derive(Copy, Clone)]
-      pub struct u16x8 { neon : uint16x8_t }
+      pub struct u16x8 { pub(crate) neon : uint16x8_t }
 
       impl Default for u16x8 {
         #[inline]
@@ -51,7 +51,7 @@ pick! {
   } else {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(16))]
-    pub struct u16x8 { arr: [u16;8] }
+    pub struct u16x8 { pub(crate) arr: [u16;8] }
   }
 }
 

diff --git a/src/u32x4_.rs b/src/u32x4_.rs
@@ -4,13 +4,13 @@ pick! {
   if #[cfg(target_feature="sse2")] {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(16))]
-    pub struct u32x4 { sse: m128i }
+    pub struct u32x4 { pub(crate) sse: m128i }
   } else if #[cfg(target_feature="simd128")] {
     use core::arch::wasm32::*;
 
     #[derive(Clone, Copy)]
     #[repr(transparent)]
-    pub struct u32x4 { simd: v128 }
+    pub struct u32x4 { pub(crate) simd: v128 }
 
     impl Default for u32x4 {
       fn default() -> Self {
@@ -29,7 +29,7 @@ pick! {
     use core::arch::aarch64::*;
     #[repr(C)]
     #[derive(Copy, Clone)]
-    pub struct u32x4 { neon : uint32x4_t }
+    pub struct u32x4 { pub(crate) neon : uint32x4_t }
 
     impl Default for u32x4 {
       #[inline]

diff --git a/src/u32x8_.rs b/src/u32x8_.rs
@@ -4,11 +4,11 @@ pick! {
   if #[cfg(target_feature="avx2")] {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(32))]
-    pub struct u32x8 { avx2: m256i }
+    pub struct u32x8 { pub(crate) avx2: m256i }
   } else {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(32))]
-    pub struct u32x8 { a : u32x4, b : u32x4 }
+    pub struct u32x8 { pub(crate) a : u32x4, pub(crate) b : u32x4 }
   }
 }
 

diff --git a/src/u64x2_.rs b/src/u64x2_.rs
@@ -4,13 +4,13 @@ pick! {
   if #[cfg(target_feature="sse2")] {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(16))]
-    pub struct u64x2 { sse: m128i }
+    pub struct u64x2 { pub(crate) sse: m128i }
   } else if #[cfg(target_feature="simd128")] {
     use core::arch::wasm32::*;
 
     #[derive(Clone, Copy)]
     #[repr(transparent)]
-    pub struct u64x2 { simd: v128 }
+    pub struct u64x2 { pub(crate) simd: v128 }
 
     impl Default for u64x2 {
       fn default() -> Self {
@@ -29,7 +29,7 @@ pick! {
     use core::arch::aarch64::*;
     #[repr(C)]
     #[derive(Copy, Clone)]
-    pub struct u64x2 { neon : uint64x2_t }
+    pub struct u64x2 { pub(crate) neon : uint64x2_t }
 
     impl Default for u64x2 {
       #[inline]

diff --git a/src/u64x4_.rs b/src/u64x4_.rs
@@ -4,11 +4,11 @@ pick! {
   if #[cfg(target_feature="avx2")] {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(32))]
-    pub struct u64x4 { avx2: m256i }
+    pub struct u64x4 { pub(crate) avx2: m256i }
   } else {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(32))]
-    pub struct u64x4 { a : u64x2, b : u64x2 }
+    pub struct u64x4 { pub(crate) a : u64x2, pub(crate) b : u64x2 }
   }
 }
 

diff --git a/src/u8x16_.rs b/src/u8x16_.rs
@@ -10,7 +10,7 @@ pick! {
 
     #[derive(Clone, Copy)]
     #[repr(transparent)]
-    pub struct u8x16 { simd: v128 }
+    pub struct u8x16 { pub(crate) simd: v128 }
 
     impl Default for u8x16 {
       fn default() -> Self {
@@ -29,7 +29,7 @@ pick! {
     use core::arch::aarch64::*;
     #[repr(C)]
     #[derive(Copy, Clone)]
-    pub struct u8x16 { neon : uint8x16_t }
+    pub struct u8x16 { pub(crate) neon : uint8x16_t }
 
     impl Default for u8x16 {
       #[inline]
@@ -51,7 +51,7 @@ pick! {
   } else {
     #[derive(Default, Clone, Copy, PartialEq, Eq)]
     #[repr(C, align(16))]
-    pub struct u8x16 { arr: [u8;16] }
+    pub struct u8x16 { pub(crate) arr: [u8;16] }
   }
 }